TWI647658B - Device, system and method for automatically identifying image features - Google Patents

Abstract

The invention discloses an image feature automatic identification device, system and method. The image feature automatic identification method comprises the following steps: performing image processing procedures on a plurality of sample images respectively having different image features, and respectively superimposing the image processing results generated by the image processing program with the standard images to generate amplification. a plurality of sample images; providing a plurality of sample images before and after amplification to a deep learning system to establish an image feature automatic recognition algorithm, including an identification standard for image features; and extracting the object to be detected The image to be tested is analyzed by the image feature automatic recognition algorithm, and the image to be tested is determined according to the identification standard.

Description

Image feature automatic identification device, system and method

The present invention relates to an image feature automatic identification device, system and method, and more particularly to an image feature automatic identification device, system and method for amplifying a sample image for a deep learning system.

With the maturity of Machine Learning technology, there are more diverse applications in image recognition, speech recognition or natural language processing.

Among them, the combination of image recognition technology and machine learning technology is particularly excellent. In addition to basic handwriting recognition, object recognition and face recognition applications, deep learning and image recognition integration technologies are often combined with automatic Automated Optical Inspection (AOI) system, which is applied to the quality control of products in the production process.

The machine learning algorithm is a kind of algorithm that automatically analyzes and obtains the law from the data and uses the law to predict the unknown data. Obviously, in order to ensure that the algorithm generated by deep learning can correctly discriminate unknown data, a large amount of sample data must be provided in the deep learning training process, in particular, a large number of labeled sample materials. So that the model of deep learning can fully and correctly learn the key to the discrimination.

However, the collection of a large number of samples is not an easy task. Especially when the deep learning system is applied to the quality control of the production process, it is usually necessary to learn the identification mark. The characteristics of the product, if the producer of the product is required to produce a very large number of tantalum products, and manually select a large number of tantalum samples and mark them, then the deep learning system can be effectively applied to help generate an algorithm that can automatically identify the characteristics of the flaw. Furthermore, by automatically recognizing the algorithm of the feature, it is obviously not an ideal solution to help the producer to carry out quality control and improve the process.

The technical problem to be solved by the present invention is to provide an automatic recognition device, system and method for image features according to the deficiencies of the prior art, which can effectively increase the number of sample images and effectively improve the training process under the deep learning system. The diversity of data can further improve the efficiency and correctness of the deep learning system applied to the automatic identification of image features.

In order to solve the above technical problem, one of the technical solutions adopted by the present invention is to provide an image feature automatic identification device, which includes a storage unit, a processing unit, and an image capturing unit, wherein the storage unit stores a database, the database storing a first image feature category group and at least one standard image, the first image feature category group storing a plurality of first sample images, each of the first The product images respectively have different first image features; the processing unit is connected to the storage unit signal; the image capturing unit is connected to the processing unit signal for capturing a to-be-detected object Measure the image. The processing unit reads a plurality of the first sample images in the database, and the plurality of first sample images perform an image processing program to respectively generate a plurality of image processing results, and the plurality of The image processing results are respectively superimposed with a plurality of the standard images to respectively generate a plurality of amplified first sample images. The processing unit executes a training program of the deep learning system according to the plurality of the first sample images before and after the amplification to establish an image feature automatic identification algorithm, and the image feature automatic identification algorithm A first identification criterion for the first image feature is included. The processing unit acquires the image to be tested from the image capturing unit, and the processing unit performs the image feature automatic identification algorithm to analyze the image to be tested, and the processing The unit determines, according to the first identification criterion, whether the image to be tested has the first image feature.

In order to solve the above technical problem, another technical solution adopted by the present invention is to provide an image feature automatic identification system, which includes a server end and a detecting end, wherein the server end includes a storage unit and a processing unit. The storage unit stores a database, the database stores a first image feature category group and at least one standard image, and the first image feature category group stores a plurality of first sample images. Each of the first sample images has a different first image feature; the processing unit is coupled to the storage unit signal. The processing unit reads a plurality of the first sample images in the database, and respectively performs an image processing program, and respectively respectively respectively respectively the image processing results generated by the image processing program The standard image is superimposed to generate a plurality of amplified first sample images; the processing unit performs a training program of the deep learning system according to the plurality of first sample images before and after the amplification To establish an image feature automatic identification algorithm, the image feature automatic recognition algorithm includes a first identification criterion for the first image feature. The detecting end is connected to the servo end signal, and the image end automatic recognition algorithm is received by the server end, the detecting end includes an image capturing module and a processing module, and the image is The module is configured to capture a to-be-tested image of the object to be detected; the processing module is coupled to the image capturing module to obtain the image to be tested from the image capturing module The processing module performs the image feature automatic identification algorithm to analyze the image to be tested, and determines whether the image to be tested has the first image feature according to the first identification criterion.

In order to solve the above technical problem, another technical solution adopted by the present invention is to provide an image feature automatic identification method, which includes the following steps: a plurality of first sample images respectively having different first image features Performing an image processing program, and respectively superimposing the image processing result generated by the image processing program with a standard image not having the first image feature to generate a plurality of amplified images a sample image; providing a plurality of first sample images before and after the amplification to a depth learning system to establish an image feature automatic recognition algorithm, wherein the image feature automatic recognition algorithm includes a first identification criterion of the first image feature; capturing a to-be-tested image of the object to be detected, and analyzing the image to be tested by the image feature automatic recognition algorithm, and according to the An identification criterion determines whether the image to be tested has the first image feature.

The invention has the beneficial effects of the image feature automatic identification device, system and method provided by the technical solution of the present invention, which can perform image processing procedures on a plurality of sample images respectively having different image features, and Technical features of superimposing standard images of image features to produce amplified multiple sample images" and "providing pre-amplification and post-amplification multiple sample images to a deep learning system" to enhance depth The diversity of data that can be accessed by the training process under the learning system can improve the efficiency and correctness of the deep learning system applied to the automatic identification of image features, and the depth can be adopted at the stage where a large number of samples are not accumulated. The learning system generates an automatic recognition algorithm for image features. In the early stage of the new product process, the image feature automatic identification technology can be effectively used to improve the process reduction and save a lot of time and cost.

For a better understanding of the features and technical aspects of the present invention, reference should be made to the accompanying drawings.

1‧‧‧ device

11, 211‧‧‧ storage unit

111, 2111‧‧ ‧ database

12.212‧‧‧Processing unit

13‧‧‧Image capture unit

2‧‧‧System

21‧‧‧Server

22‧‧‧Detection

221‧‧‧Processing module

222‧‧‧Image capture module

23‧‧‧ Sample image supply end

P0‧‧‧ standard image

D1‧‧‧ first image features

P11...P1n, P1(n+1)...P1N, P111, P112, P113‧‧‧ first sample image

G1‧‧‧First Image Feature Category Group

D2‧‧‧Second image features

P21~P2N, P211, P212‧‧‧ second sample image

G2‧‧‧Second image feature category group

1 is a functional block diagram of an automatic image feature recognition apparatus according to a first embodiment of the present invention.

2A is a schematic diagram showing an adjustment result of adjusting an image shape of a first sample image by the image processing program according to the first embodiment of the present invention.

FIG. 2B is a schematic diagram showing an adjustment result of adjusting the gray scale of the image of the second sample image by the image processing program according to the first embodiment of the present invention.

3 is a functional block of an image feature automatic identification system according to a second embodiment of the present invention; Figure.

FIG. 4 is a main flowchart of a method for automatically identifying an image feature according to a third embodiment of the present invention.

FIG. 5 is a flowchart of a pre-verification procedure performed by an automatic image feature recognition method according to a fourth embodiment of the present invention.

The embodiments of the present invention relating to the "image feature automatic identification device, system and method" are described by way of specific embodiments, and those skilled in the art can understand the advantages and effects of the present invention from the contents disclosed in the specification. The present invention may be carried out or applied in various other specific embodiments, and various modifications and changes can be made without departing from the spirit and scope of the invention. In addition, the drawings of the present invention are merely illustrative and are not intended to be stated in the actual size. The following embodiments will further explain the related technical content of the present invention, but the disclosure is not intended to limit the scope of the present invention.

[First Embodiment]

Referring to FIG. 1, FIG. 2A and FIG. 2B, FIG. 1 is a functional block diagram of an automatic image feature recognition apparatus according to a first embodiment of the present invention. 2A is a schematic diagram showing an adjustment result of adjusting an image shape of a first sample image by the image processing program according to the first embodiment of the present invention. FIG. 2B is a schematic diagram showing an adjustment result of adjusting the gray scale of the image of the second sample image by the image processing program according to the first embodiment of the present invention. As shown in the above figure, the first embodiment of the present invention provides an image feature automatic identification device 1 including a storage unit 11, a processing unit 12, and an image capturing unit 13.

Referring to FIG. 1 , the storage unit 11 stores a database 111. The database 111 stores a first image feature category group G1, a second image feature category group G2, and at least one standard image P0. The first image feature category group G1 stores a plurality of first sample images P11~P1N; the second image feature category group G2 stores a plurality of second Sample image P21~P2N. The processing unit 12 is connected to the storage unit 11 and the image capturing unit 13 is connected to the processing unit 12 for capturing the image to be detected (not shown).

For convenience of description, the operation of the automatic image recognition apparatus 1 of the first embodiment of the present invention will be described below by taking the detection of the flaw of the contact lens as an example. Please refer to the function block diagram of FIG. 1 together with FIG. 2A and FIG. 2B. In the first embodiment of the present invention, the storage unit 11 of the device 1 stores a contact lens image (ie, a standard image P0) that is deemed to be qualified after manual screening, and a plurality of different contact lens images (ie, the first The sample images P11 to P1N and the second sample images P21 to P2N), the aforementioned acceptable contact lens images and a plurality of different contact lens images are stored in the database 111.

In this embodiment, the plurality of first sample images P11~P1N respectively have different first image features D1, and the plurality of second sample images P21~P2N respectively have different second image features D2. More specifically, when a specific contact lens has a bubble, the image of the sample obtained by photographing the contact lens has a "bubble" characteristic compared to a qualified contact lens image (hereinafter referred to as "bubble" In this embodiment, the 瑕疵 feature of the "bubble 瑕疵" is taken as the first image feature D1, and the sample image having the first image feature D1 is the first sample image in this embodiment. P11~P1N. According to the captured contact lens, the first image feature D1 (i.e., bubble) on each of the first sample images P11 to P1N is slightly different. On the other hand, when a specific contact lens is produced during the production process, a flaw is generated at the edge during demolding, and the image of the sample obtained by photographing the contact lens has an "edge" compared to a qualified contact lens image. In this embodiment, the 脱 feature of the "release 瑕疵" is used as the second image feature D2, and the sample having the second image feature D2 is produced. The image is the second sample image P21 to P2N in the present embodiment. According to the captured contact lens, the second image feature D2 (i.e., "release mold") on each of the second sample images P21 to P2N is also slightly different.

In order for the deep learning system to correctly identify different defects, the features must be properly classified and labeled. In addition, it is more important to provide a sufficient number of sample images for deep learning systems to learn. In this embodiment, the manufacturer of the contact lens may first manually select a sample image of the flaw in an artificial manner (of course, with the aid of the AOI detection device), and according to the type of the flaw, it will be classified according to the type of the flaw. The first sample images P11~P1N of an image feature D1 are stored in the first image feature category group G1 of the database 111, and the second sample images P21~P2N having the second image feature D2 are stored in the database. The second image feature category group G2 of 111. For example, 100 pieces of first sample image P11~P1N data having the first image feature D1 and 100 pieces of second sample image P21~P2N data having the second image feature D2 may be separately stored in the database 111. in.

However, in order to enable the subsequent generated algorithm to accurately identify the image features automatically, the present invention first amplifies the sample image. In this embodiment, the processing unit 12 first reads a plurality of sample images in the database 111, and performs image processing procedures on the plurality of sample images to respectively generate a plurality of image processing results. The image processing program includes one or a combination of two or more of an image shape adjustment program, an image contrast adjustment program, an image grayscale adjustment program, and an image color temperature adjustment program. The processing unit 12 superimposes the plurality of image processing results on the plurality of standard images P0 to generate a plurality of amplified sample images, respectively.

For example, as shown in FIG. 2A , in the embodiment, the processing unit 12 first reads the first sample image P11 in the database 111 and processes the first sample through the image shape adjustment program. The image P11 is stretched or deformed to generate a plurality of image processing results, and the plurality of image processing results are respectively superimposed with the standard image P0 to respectively generate the amplified plurality of first sample images P111 to P113. . More specifically, the first image feature D1 of the first sample image P11 (ie, the "bubble" portion of the circle that is circled on the image) is stretched or deformed, and the image processing results after the deformation are respectively It is superimposed with the standard image P0.

Referring to FIG. 2B, in the embodiment, the processing unit 12 can also be used first. Reading the second sample image P21 in the database 111 and the second image feature D2 of the second sample image P21 by the image grayscale adjustment program (or only for the image on the edge of the contact lens) The grayscale adjustment is performed to generate a plurality of image processing results, and the plurality of image processing results are respectively superimposed with the standard image P0 to respectively generate the plurality of amplified second sample images P211 and P212.

Although the first sample image P11 is stretched or deformed by the image shape adjustment program, and the grayscale adjustment is performed on the second sample image P21 by the image grayscale adjustment program, the type of the image processing program and the first There is no fixed correspondence between the categories of an image feature D1 or the second image feature D2. In addition, the image shape adjustment program, the image contrast adjustment program, the image grayscale adjustment program, and the image color temperature adjustment program are not mutually exclusive relationships, but can be combined with each other for image adjustment, and even other image adjustment methods are added. The sample image is adjusted to increase the diversity of image processing results. According to this, it is possible to amplify the first sample images P11 to P1N and the second sample images P21 to P2N, which originally had only 100 pens, in a large amount (for example, to 100,000 pieces of data, respectively).

It should be particularly noted that, in this embodiment, the database 111 stores only one standard image P0, but the present invention is not limited thereto, and specific methods may also be used in each image feature category group. , respectively, store a standard image P0, and perform image processing and overlay operations, respectively.

Next, the processing unit 12 performs a training program of the deep learning system to establish image features according to the plurality of first sample images P11 P P1N and the plurality of second sample images P21 P P2N before and after the amplification. Automatic identification algorithm. Specifically, the deep learning system used may be a framework of Caffe, Theano, TensorFlow, or Lasagne, Keras, or even DSSTNE, and the present invention does not specifically limit which framework to use. The image feature automatic recognition algorithm generated by the foregoing method includes a first identification criterion for the first image feature D1 ("bubble" in the embodiment) and a second image feature D2 (in the embodiment) The second identification standard for "demolition".

After the image feature automatic recognition algorithm is established, the device 1 of the present invention can be detected. Specifically, the image capturing unit 13 (such as an image capturing lens) captures the image to be detected of the object to be detected, and then the processing unit 12 obtains the image to be tested from the image capturing unit 13 and performs an automatic image feature recognition algorithm. Analyze the image to be measured. The processing unit 12 determines whether the image to be tested has the first image feature D1 according to the first identification criterion, and determines whether the image to be tested has the second image feature D2 according to the second identification criterion.

So far, it is the basic operation mode of the image feature automatic identification device 1 of the present invention. In this embodiment, by amplifying the first sample images P11~P1N and the second sample images P21~P2N in a large amount, a more accurate image feature automatic recognition calculation can be established in the training program of the deep learning system. law.

[Discovery and establishment of 瑕疵 categories]

It should be particularly noted that although the above description is made by taking "bubble" and "release" as an example, the present invention is not limited to setting a plurality of types of defects for detection at the beginning, and may be set only at the beginning. A defect feature is detected, and in the process of performing the image feature automatic recognition algorithm to analyze the image, the other feature categories to be tested are selected and established according to the detection result.

For example, it is assumed that only the first image feature category group G1 is stored in the original database 111, and the generated image feature automatic recognition algorithm only sets the first image feature of the first sample image P11~P1N. D1 is the first identification criterion for identification. However, in the process of verifying the plurality of images to be tested by the image feature automatic recognition algorithm, the processing unit 12 determines that the plurality of images to be tested do not conform to the standard image P0, but both have the second image feature D2. The processing unit 12 is configured to create a second image feature D2 category group G2 in the database 111, and record each of the plurality of image to be tested having the second image feature D2 as a plurality of second sample images P21, respectively. And stored in the second image feature D2 category group G2. In this way, it is helpful to discover and establish new categories of defects so that they can be more proactive in making immediate improvements to unknown defects on the product.

[Verification rate verification and improvement]

Through the flow described above, it is sufficient that the deep learning system of the present invention can perform training using a very large and diverse sample image, and can generate an image feature automatic recognition algorithm with relatively accurate judgment. However, the sample image generated after the amplification may of course be subjected to the image processing program again, and further amplified into a larger number of sample images, which may further improve the accuracy of the image feature automatic recognition algorithm recognition. In order to confirm whether it is necessary to amplify the sample image again, the device 1 of the present invention can pre-verify the generated image feature automatic identification algorithm before actually using the image feature automatic recognition algorithm to test the image to be tested. Identify the correct rate of the job.

Specifically, before processing the training program of a deep learning system, the processing unit 12 firstly converts a part (for example, 20%) of the first sample images P11~P1n of the plurality of first sample images P11~P1N. The image for the first verification is selected, and the training program of the deep learning system is executed with the remaining plurality (for example, the remaining 80%) of the first sample images P1(n+1) to P1N. Moreover, after the image learning automatic identification algorithm is established in the deep learning system, the processing unit 12 performs an image feature automatic identification algorithm to verify the previously selected first verification image. That is to say, according to the first identification criterion in the image feature automatic recognition algorithm, it is determined whether the first verification image has the first image feature D1 to confirm the correctness of the first identification standard in the image feature automatic recognition algorithm. . If it is not recognized that the first verification image has the first image feature D1, or the image that does not have the first image feature D1 (eg, the standard image P0) is identified as having the first image feature D1, the recognition result is incorrect. Conversely, the recognition result is correct.

The user of the device 1 can preset the first correct rate threshold by the processing unit 12. In addition, the number of the first verification images can be adjusted as needed. The ratio occupied in the first sample image P11. When actually performed, the number of images for the first verification can be set to 3% to 50% in all the first sample images P11. After performing the image feature automatic identification algorithm, and determining whether the standard image P0 and the plurality of first verification images have the first image feature D1 according to the first identification criterion, the recognition result is recorded correctly or incorrectly, in other words, The first identification correctness rate of the first image feature D1 is recorded according to the determination result.

Next, the first identification accuracy rate is compared with a preset first correct rate threshold value. When the first recognition correct rate is lower than the first correct rate threshold, it may not be sufficient to indicate that the sample image provided for training by the deep learning system is performed. Therefore, in order to be able to perform further training, the processing unit 12 will have multiple A sample image P11 is again subjected to an image processing program to further generate a plurality of amplified first sample images P11. Of course, in order to verify the sample image after re-amplification, it is provided to the deep learning system for training, and after correcting the first identification standard, whether it has been able to generate an accurate image feature automatic recognition algorithm, therefore, still has to be performed The aforementioned verification procedure. Specifically, the processing unit 12 selects 3% to 50% of the plurality of first sample images P11 after re-amplification as the first verification image, and the remaining plurality of first sample images The P11 is again provided to the deep learning system to correct the first identification criterion, and the corrected first identification criterion is used to determine again whether the standard image P0 and the plurality of first verification images have the first image feature D1, and again Obtain the first identification correct rate. That is to say, if the first correct rate threshold cannot be reached, it will return to the image processing program to further amplify the sample image. The image feature automatic recognition algorithm is actually applied to the detection of the object to be detected until the first recognition accuracy rate is higher than (including just reaching) the first correct rate threshold, and the image feature automatic recognition algorithm is executed by the processing unit 12. And determining, according to the first identification criterion that is finally corrected in the algorithm according to the image feature, determining whether the image to be detected extracted from the object to be detected has the first image feature D1. In order to avoid setting the threshold of the first correct rate at the beginning to be too high, the standard is not up to standard, and the image processing program continuously enlarges the sample image, and other stop conditions can be set, but it is not emphasized by the present invention. Heavy Point, how to set the specific conditions is not described here.

[pre-test]

On the other hand, one problem that needs to be considered is that if the diversity of the first image feature D1 is insufficient in the sample image originally provided, it is likely to affect the subsequent image feature automatic recognition algorithm being performed. The correctness of the recognition, even if the sample image is repeatedly amplified, it may be impossible to carry out an efficient training program because the diversity at the beginning is too low. In order to solve this problem, in the present embodiment, the apparatus 1 can also cause the processing unit 12 to execute a pre-verification program in advance before executing the image processing program.

Specifically, the processing unit 12 selects 3% to 50% of the plurality of first sample images P11 to P1N (which have not been subjected to any image processing program) as the first pre-examination image, and the first sample The remaining plurality of first sample images P11~P1N in the image P11 are provided to the depth learning system to establish a first pre-test standard including the first image feature D1. In this process, since it is only necessary to initially confirm whether the obtained first sample images P11 to P1N have sufficient diversity, it is possible to adopt a deep learning system with a simplified structure. Likewise, the threshold is also set by the processing unit 12, which is defined herein as the first pre-test threshold. Since only a preliminary judgment is made on the diversity of the first sample images P11 to P1N, the first pre-test threshold value does not have to be set too high. Compared with the foregoing first correct rate threshold for ensuring the final judgment accuracy, the value of the first pre-test threshold value here is required to be low, and therefore, in the case where both the pre-test and the correct rate verification are taken, The first pre-test threshold value will be lower than the first correct rate threshold. The processing unit 12 determines whether the standard image P0 and the plurality of first verification images have the first image feature D1 according to the first pre-verification standard, and records the first pre-test standard for the first image feature D1 according to the determination result. Pre-test the correct rate.

In the foregoing pre-test procedure, if the first pre-test correct rate has reached the first pre-test threshold value, it indicates that the diversity of the first sample image P11~P1N has reached the desired level. Standard, so the main flow of the present invention can be continued, in other words, the image processing program can be executed next; otherwise, if the first pre-test correct rate is lower than the first pre-test threshold, the first sample image P11 is indicated. The diversity of ~P1N is actually not enough. At this time, the program should be terminated. Do not perform subsequent image processing procedures. Instead, first collect more first sample images P11~P1N with different first image features D1. To ensure that the deep learning system can be effectively and effectively trained.

[Second embodiment]

Referring to FIG. 3, FIG. 3 is a functional block diagram of an image feature automatic identification system 2 according to a second embodiment of the present invention. As can be seen from the above figure, the second embodiment of the present invention provides an image feature automatic identification system 2, which includes a servo end 21, a detecting end 22, and a sample image supply end 23, and the sample image supply end 23 is also connected to the servo end 21 signal. Connect to provide. In this embodiment, the server 21 is a provider of the image feature automatic identification algorithm, and the detecting end 22 is connected to the servo end 21 signal, and the image end automatic recognition algorithm can be received by the servo end 21, so that the detecting end 22 can The image feature automatic identification algorithm is executed, and the image to be tested of the object to be detected is detected and analyzed. As for the sample image supply end 23, it is also connected to the servo end 21 signal, and a plurality of first sample images P11 to P1N can be supplied to the servo end 21. In the actual architecture of system 2, detection terminal 22 may be identical to sample image supply 23, but may be separately provided.

In particular, as can be seen from the block diagram shown in FIG. 3, in the present embodiment, the servo terminal 21, the detecting terminal 22, and the sample image supply terminal 23 are connected to each other via a network, and can exchange signals with each other. (for example, transmitting a sample image or an image feature automatic recognition algorithm), however, the "signal connection" referred to in the present invention is not limited thereto, and the servo terminal 21, the detection terminal 22, and the sample image supply terminal 23 are at either end. The data is stored in a storage medium such as a compact disc, a flash memory or a hard disk, and the data is provided to any other end. It also conforms to the "signal connection" referred to herein, and is hereby indicated.

Similarly, the detection of the contact lens is taken as an example. In short, the sample image supply end 23 (contact lens production and tester) can automatically recognize the "bubble" in order to obtain 影像" (first image feature D1) image feature automatic recognition algorithm, so a plurality of first sample images P11~P1N having different first image features D1 are supplied to the server 21 (algorithm provider) . After receiving the plurality of first sample images P11~P1N, the servo end 21 generates an image feature automatic identification algorithm through a series of steps, and provides an image feature automatic identification algorithm to the detecting end 22 (contact lens production and tester) So that the detecting end 22 can apply the image feature automatic recognition algorithm to the detection of the produced contact lens (to be detected). In other words, although the servo end 21 does not perform the detection of the contact lens, the material for generating the image feature automatic recognition algorithm is obtained from the sample image supply terminal 23 where it is required, and the finished product (image feature automatic recognition algorithm) Provided to the detection terminal 22 which is also in need and will perform an image feature automatic recognition algorithm.

Specifically, the server 21 includes a storage unit 211 and a processing unit 212. The storage unit 211 stores a database 2111. The database 2111 stores a first image feature category group G1 and at least one standard image P0. The category group G1 stores a plurality of first sample images P11 to P1N (provided by the sample image supply terminal 23), and each of the first sample images P11 to P1N has a different first image feature D1. The processing unit 212 is connected to the storage unit 211, and can read the plurality of first sample images P11~P1N in the database 2111, and respectively perform an image processing program. The image processing program includes one or a combination of two or more of an image shape adjustment program, an image contrast adjustment program, an image grayscale adjustment program, and an image color temperature adjustment program. The processing unit 212 respectively superimposes the image processing results generated by the image processing program with the standard image P0 to generate a plurality of amplified first sample images P11 to P1N.

The processing unit 212 performs a training program of the deep learning system according to the plurality of first sample images P11~P1N before and after the amplification to establish an automatic feature recognition algorithm for image features, and the image feature automatic recognition algorithm includes The first identification criterion of the first image feature D1.

More specifically, in the present embodiment, the processing unit 212 is performing deep learning. Before the training program of the system, at least one of the plurality of first sample images P11~P1N is selected as the first verification image, and the remaining plurality of first sample images P11 are selected. ~P1N performs the training program of the deep learning system. In actual operation, the number of first verification images is plural, and the number of first verification images accounts for 3% to 50% of the number of the plurality of first sample images P11 to P1N. In addition, the processing unit 212 sets a first correct rate threshold, and after the depth learning system establishes an image feature automatic identification algorithm, the processing unit 212 performs an image feature automatic identification algorithm to determine the standard image P0 according to the first identification criterion. Whether the plurality of first verification images have the first image feature D1 to confirm the correctness of the first identification criterion. The processing unit 212 records the first recognition accuracy rate of the first image feature D1 by the first identification criterion according to the determination result, and compares the first identification accuracy rate with the foregoing first correct rate threshold value.

When the first identification correct rate is lower than the first correct rate threshold, the processing unit 212 performs the image processing procedure on the plurality of first sample images P11~P1N to further generate the amplified plurality of first sample images. Like P11~P1N, and from the plurality of first sample images P11 to P1N after re-amplification, 3% to 50% are again selected as the first verification image, and the remaining plurality of first samples are selected. The images P11~P1N are again provided to the deep learning system to correct the first identification criterion, and the first identification standard is corrected to determine whether the standard image P0 and the plurality of first verification images have the first image feature D1. And get the first recognition correct rate again. On the other hand, if the first identification correct rate has reached the first correct rate threshold, the image feature automatic recognition algorithm can be provided to the detecting end 22.

The detecting end 22 includes an image capturing module 222 and a processing module 221, and the image capturing module 222 is configured to capture an image to be detected of the object to be detected. The processing module 221 is connected to the image capturing module 222, and the image to be tested can be obtained from the image capturing module 222. After the detecting end 22 receives the image feature automatic identification algorithm from the server 21, the processing module 221 performs an image feature automatic identification algorithm to analyze the image to be measured, and determines whether the image to be tested has the first image according to the first identification criterion. Feature D1.

In particular, since the database 2111 in this embodiment is located The servo terminal 21 is not at the detecting end 22. Therefore, in the embodiment, when the processing module 221 of the detecting terminal 22 determines that the plurality of images to be tested do not conform to the standard image P0, but both have the second image feature D2, The detecting end 22 respectively records a plurality of images to be tested having the second image feature D2 as a plurality of second sample images, and supplies the plurality of second sample images to the servo end 21. After that, the processing unit 212 of the server 21 creates a second image feature category group G2 in the database 2111, and stores a plurality of second sample images received from the detecting terminal 22 in the second image feature category. In group G2.

After the new category is established, the processing unit 212 of the server 21 can further read the plurality of second sample images in the database 2111, and respectively perform image processing procedures, and separately generate images generated by the image processing program. The processing results are each superimposed with the standard image P0 to produce a plurality of amplified second sample images. And the processing unit 212 executes the training program of the deep learning system according to the plurality of first sample images P11~P1N before and after the amplification to establish the second image feature D2 in the image feature automatic recognition algorithm. Two identification criteria. After completion, the server 21 can re-provide the image feature automatic recognition algorithm to the detecting end 22. After the detecting end 22 receives the image feature automatic recognition algorithm again, and the processing module 221 analyzes the image to be measured by the image feature automatic identification algorithm, it determines whether the image to be tested has the first according to the first identification criterion. An image feature D1, and determining whether the image to be tested has the second image feature D2 according to the second identification criterion.

As described above, the present invention enables the detecting end 22 to immediately feed back the actual use result of the image feature automatic recognition algorithm to the servo terminal 21, so that the servo terminal 21 can perform further analysis.

In another embodiment, after the detecting end 22 uses the image feature automatic identification algorithm to detect, it is found that a defective product is not detected (whether or not the 瑕疵 feature to be recognized is set in advance), and the manual recording method corresponds to The image to be tested is stored as a sample image and supplied to the servo terminal 21 for analysis. If the server 21 finds a new category after the analysis, the detection terminal 22 is separately notified to adjust and optimize the process to form a benign interaction process.

As mentioned in the first embodiment, if the diversity of the first image feature D1 is insufficient in the initially provided sample image, it is likely to affect the subsequent image feature automatic recognition algorithm for recognition. The correctness of time. Therefore, in the present embodiment, the apparatus 1 can also cause the processing unit 12 to execute a pre-verification program in advance before executing the image processing program. After the servo terminal 21 receives the plurality of first sample images P11 to P1N from the sample image supply terminal 23, it is also possible to perform the inspection program to confirm whether the received first sample images P11 to P1N have sufficient. Diversity. Specifically, the processing unit 212 of the servo terminal 21 selects 3% to 50% (preferably 15% to 25%) of the plurality of first sample images P11 to P1N as the first pre-inspection image, And the remaining plurality of first sample images P11 to P1N in the first sample images P11 to P1N (ie, 50% to 97% of the first sample images P11 to P1N, preferably 75% to 85) The % first sample images P11~P1N) are provided to the deep learning system to establish a first pre-test standard including the first image feature D1. Similarly, the first pre-check threshold is also set by the processing unit 212 in this embodiment. The processing unit 212 determines whether the standard image P0 and the plurality of first verification images have the first image feature D1 according to the first pre-test standard established by the depth learning system, and records the first pre-test standard against the first image according to the determination result. The first pre-test correct rate of feature D1. After obtaining the first pre-test correct rate, the processing unit 212 compares the first pre-test correct rate with a preset first pre-test threshold value.

If the first pre-test correct rate is lower than the first pre-test threshold, it means that the first sample image P11~P1N provided by the sample image supply end 23 is insufficient in diversity. At this time, in this embodiment, the servo end 21 does not perform the subsequent movie item amplification process, but requests the sample image supply terminal 23 for more different first sample images P11 to P1N to enrich the diversity of the first sample images P11 to P1N. Of course, after receiving more first sample images P11~P1N, it is still necessary to confirm again whether the diversity has reached the standard by the inspection program. Conversely, when the first pre-test correct rate reaches the first pre-check threshold, the servo 21 can continue the main flow of the present invention. In other words, the first sample image P11 will be executed. P1N image processing program For amplification.

[Third embodiment]

Please refer to Figure 4. FIG. 4 is a main flowchart of a method for automatically identifying an image feature according to a third embodiment of the present invention. The main flow of the automatic image feature recognition method provided by the present invention will be described below with reference to FIG. The image feature automatic identification method of the present invention mainly comprises the following steps: S100: obtaining a sample image; S102: amplifying the sample image by the image processing program; S104: selecting a verification image in the plurality of sample images; S106: The remaining sample images are provided for training in the deep learning system; S108: verifying the image feature automatic recognition algorithm generated by the depth learning system by the verification image; S110: determining whether the recognition correct rate reaches the correct rate threshold, and if so, entering the step S112; if not (below), return to step S102; S112: extract the image to be detected of the object to be detected; S114: identify the image to be tested by the image feature automatic recognition algorithm.

Specifically, after acquiring the sample image, the image feature automatic identification method of the present invention first performs an image processing program for the sample image to perform amplification (steps S100 and S102). The specific method is to perform one or two of an image shape adjustment program, an image contrast adjustment program, an image gray scale adjustment program, and an image color temperature adjustment program for a plurality of first sample images having different first image features. The above combination is performed, and the image processing result generated by the image processing program is superimposed with the standard image without the first image feature to generate a plurality of amplified first sample images.

In order to verify whether the image of the first sample after amplification is sufficient to establish an automatic image recognition algorithm with accurate inspection results, it is pre- At least one of the plurality of first sample images is selected as the first verification image (step S104). At the same time, the first correct rate threshold can also be set together. In the present embodiment, the number of first verification images accounts for 3% to 50% of the number of the plurality of first sample images (for example, 20% is taken). The first correct rate threshold can be adjusted according to the fault tolerance rate of the product. For example, the "bubble 瑕疵" for contact lenses is a serious flaw, and the detection difficulty is not very high. The high correct rate, at this time, the first correct rate threshold can be set to 99% or more. On the other hand, when the image feature automatic identification method of the present invention is simultaneously applied to detect the second image feature, the second verification image may be selected for the second image feature and the second sample image having the second image feature. Like, and set the second correct rate threshold. For the "release mold" of the contact lens, since it is difficult to accurately determine the flaw, if the second image feature is the "release mold" of the contact lens, the second correct rate threshold can be considered to be set to 90~95. %about.

Next, a plurality of first sample images (or a second sample image) before and after the amplification are provided to the deep learning system for training to establish an image feature automatic recognition algorithm (step S106). The image feature automatic recognition algorithm includes a first identification criterion for the first image feature.

Since a part of the plurality of first sample images has been selected as the first verification image, after the image learning automatic recognition algorithm is established in the deep learning system, the first verification map is determined according to the first identification criterion. Whether the image has the first image feature, thereby checking the image feature automatic recognition algorithm generated by the depth learning system (step S108). The specific method is: performing an image feature automatic identification algorithm, and determining, according to the first identification criterion, whether the standard image and the plurality of first verification images have the first image feature, and recording the first identification standard to the first image according to the determination result First identifying the correct rate of the feature, and then comparing the first identification correct rate with a preset first correct rate threshold to confirm whether the first identification correct rate is higher than a preset first correct rate threshold (step S110).

If the first identification correct rate is lower than the first correct rate threshold, return to the step S102: Perform a video processing procedure on the plurality of first sample images to further generate the amplified plurality of first sample images, and sequentially perform steps S104 to S110 to re-amplify the plurality of samples. In the first sample image, 3% to 50% is again selected as the first verification image, and the remaining plurality of first sample images are again supplied to the depth learning system to correct the first identification standard, and And determining, by the corrected first identification criterion, whether the standard image and the plurality of first verification images have the first image feature, and obtaining the first recognition accuracy rate again. After obtaining the first identification correct rate again, the first correct rate threshold is compared with the preset first correct rate threshold until the first identification correct rate reaches the first correct rate threshold.

After the first identification correct rate reaches the first correct rate threshold, the established image feature automatic identification algorithm can be actually applied to detect the object to be detected. Specifically, the image to be tested of the object to be detected is first captured (step S112), and then the image to be measured is analyzed by the image feature automatic identification algorithm, and the image to be tested is determined according to the first identification criterion. There is a first image feature (step S114).

The above is the main flow of the image feature automatic identification method of the present invention. Of course, the image feature automatic identification method of the present invention can also record a plurality of images to be tested as multiple seconds when determining that a plurality of images to be tested do not conform to the standard image, but both have the second image feature. Sample image. For details of this part, reference may be made to the first and second embodiments, and the details are not repeated here. In addition, when the image feature automatic identification method of the present invention is applied to a plurality of second sample images having different second image features, the following procedure is also performed: a plurality of different images having different second image features The two sample images are subjected to an image processing program, and the image processing results generated by the image processing program are superimposed with the standard image having no second image features to generate a plurality of amplified second sample images; The plurality of second sample images before and after the amplification are provided to the deep learning system to establish a second identification standard for the second image feature in the image feature automatic recognition algorithm; and the object to be detected is to be tested The image is analyzed by the image feature automatic recognition algorithm for the image to be measured; wherein, according to the first identification criterion, it is determined whether the image to be tested has the first image feature; wherein, according to the first The second identification criterion determines whether the image to be tested has a second image feature.

[Fourth embodiment]

In order to ensure that the initial sample image is sufficiently diverse, the image feature automatic identification method of the present invention may be added to the middle of step S100 to step S102 of the foregoing third embodiment, plus a pre-test process. Referring to FIG. 5, FIG. 5 is a flowchart of a pre-verification program for performing automatic image feature recognition according to a fourth embodiment of the present invention. The image feature automatic identification method of the present invention mainly comprises the following steps when performing the pre-test procedure: S100: obtaining a sample image; S116: selecting a pre-test image in a plurality of sample images; S118: selecting the remaining sample image Providing training to the deep learning system; S120: checking the sample image diversity with the image for pre-test; S122: determining whether the recognition correct rate is higher than the pre-test threshold value, and if yes, proceeding to step S102; if not (below), Going back to step S100; S102: amplifying the sample image by the image processing program; specifically, the image feature automatic identification method of the present invention firstly takes a plurality of first sample images after obtaining the sample image 3% to 50% is selected as the first pre-examination image (step S100 and step S116). At the same time, the first pre-test threshold value can also be set together. As mentioned before, the first pre-test threshold value does not have to be as high as the first correct rate threshold, so it can be set to 70% to 90%. Of course, the specific situation can be adjusted as needed.

Next, the remaining plurality of first sample images in the first sample image are provided to the deep learning system for training to establish a first pre-test criterion including the first image feature (step S118).

After the foregoing steps are completed, the sample image diversity is checked with the pre-examination image (step S120). Specifically, it is based on the first pre-test standard to judge the standard image and Whether the first pre-test image has the first image feature, and according to the determination result, the first pre-test correct rate of the first image feature is recorded by the first pre-test standard.

Next, the first pre-test correctness rate is compared with the first pre-test threshold value (step S122). Wherein, when the first pre-test correct rate is lower than the first pre-test threshold, indicating that the diversity of the first sample image is insufficient, it is necessary to increase the number of first sample images having different first image features. Therefore, returning to step S100, more first sample images having different first image features are obtained; otherwise, when the first pre-test correct rate reaches the first pre-test threshold, the image of the present invention can be successively performed. The main flow of the feature automatic identification method, in other words, proceeds to step S102, where the image processing program is executed to amplify the first sample image.

[Advantageous Effects of Embodiments]

The invention has the beneficial effects of the image feature automatic identification device, system and method provided by the technical solution of the present invention, which can perform image processing procedures on a plurality of sample images respectively having different image features, and have no image The feature image of the feature is superimposed to produce a plurality of sample images of the amplification" and the technical features of "providing a plurality of sample images before and after the amplification to the deep learning system" to enhance the deep learning system The diversity of data that can be accessed during the training process can improve the efficiency and correctness of the deep learning system for automatic recognition of image features, and the deep learning system can be used at the stage where a large number of samples are not accumulated. The image feature automatic recognition algorithm is generated. In the early stage of the new product process, the image feature automatic identification technology can be effectively used to improve the process reduction and save a lot of time and cost.

The above disclosure is only a preferred embodiment of the present invention, and is not intended to limit the scope of the present invention. Therefore, any equivalent technical changes made by using the present specification and the contents of the drawings are included in the application of the present invention. Within the scope of the patent.

Claims (19)

An image feature automatic identification device includes: a storage unit storing a database, the database storing a first image feature category group and at least one standard image, the first image feature category group storing Having a plurality of first sample images, each of the first sample images having a different first image feature; a processing unit coupled to the storage unit signal; and an image capture unit Connected to the processing unit signal for capturing an image to be detected of the object to be detected: wherein the processing unit reads a plurality of the first sample images in the database, The first sample image is subjected to an image processing program to respectively generate a plurality of image processing results, and the plurality of the image processing results are respectively superimposed with the plurality of the standard images to respectively generate a plurality of amplification images respectively a first sample image; wherein the processing unit executes a training program of the deep learning system according to the plurality of the first sample images before and after the amplification to establish an image feature automatic identification algorithm The image feature automatic recognition algorithm includes a first identification criterion for the first image feature; wherein the processing unit acquires the image to be tested from the image capturing unit, the processing unit Performing the image feature automatic identification algorithm to analyze the image to be tested, and the processing unit determines, according to the first identification criterion, whether the image to be tested has the first image feature; When the processing unit determines that the plurality of the images to be tested do not conform to the standard image, but both have a second image feature, create a second image feature category group in the database, and respectively The images to be tested having the second image feature are each recorded as a plurality of second sample images and stored in the second image feature category group. The device according to claim 1, wherein the image processing program comprises one or a combination of two or more of an image shape adjustment program, an image contrast adjustment program, an image grayscale adjustment program, and an image color temperature adjustment program. The apparatus of claim 1, wherein the processing unit further performs the program of: at least one of the plurality of the first sample images is identical before performing a training program of a deep learning system The product image is selected as a first verification image, and the training program of the deep learning system is executed with the remaining plurality of the first sample images; and the image feature is automatically established in the depth learning system After the algorithm is identified, determining whether the first verification image has the first image feature according to the first identification criterion to confirm the correctness of the first identification criterion. The device of claim 3, wherein the number of the first verification images is plural, and the number of the first verification images accounts for 3 of the plurality of first sample images. % to 50%, the processing unit further performs the following procedure: setting a first correct rate threshold; and performing the image feature automatic identification algorithm, and determining the standard image according to the first identification criterion and Whether a plurality of the first verification images have the first image feature, and recording, according to the determination result, a first recognition accuracy rate of the first image identification to the first image feature; wherein, in the Performing the following steps when the recognition correct rate is lower than the first correct rate threshold: performing the image processing procedure on the plurality of first sample images to further generate the amplified plurality of first sample images And from the plurality of first sample images after re-amplification, 3% to 50% are again selected as the first verification image, and the remaining plurality of first sample images are again extracted Supplying the depth learning system to correct the first identification criterion, and determining, by the corrected first identification criterion, whether the standard image and the plurality of first verification images have the first An image feature, and obtaining the first recognition accuracy rate again; wherein, when the first recognition accuracy rate reaches the first correct rate threshold, determining, according to the first identification criterion, extracting from the image Whether the image to be tested of the detected object has the first image feature. The device of claim 1, wherein the processing unit further performs the following steps: reading a plurality of the second sample images in the database, and performing the image processing programs separately, and respectively And superimposing the image processing results generated by the image processing program with the standard image to generate a plurality of amplified second sample images; according to the plurality of first samples before and after the amplification Performing a training program of the depth learning system to establish a second identification criterion for the second image feature in the image feature automatic recognition algorithm; and determining the waiting according to the first identification criterion Detecting whether the image has the first image feature, and determining whether the image to be tested has the second image feature according to the second identification criterion. The device of claim 1, wherein the processing unit further performs the following procedure before executing the image processing program: selecting 3% to 50% of the plurality of the first sample images as a first pre-examination image, and providing a remaining plurality of first sample images in the first sample image to the depth learning system to establish a one including the first image feature a first pre-test standard; setting a first pre-test threshold; and Determining, according to the first pre-test standard, whether the standard image and the plurality of first verification images have the first image feature, and recording the first pre-test standard to the first according to the determination result a first pre-test accuracy rate of the image feature; wherein the image processing program is executed when the first pre-test correct rate reaches the first pre-test threshold; and wherein the first pre-test When the correct rate is lower than the first pre-test threshold, the program is terminated. An image feature automatic identification system, comprising: a server, comprising: a storage unit, wherein the storage unit stores a database, the database stores a first image feature category group and at least one standard image, An image feature category group stores a plurality of first sample images, each of the first sample images respectively having different first image features; and a processing unit connected to the storage unit signal; The processing unit reads a plurality of the first sample images in the database, and respectively performs an image processing program, and respectively respectively respectively respectively the image processing results generated by the image processing program The standard image is superimposed to generate a plurality of amplified first sample images; wherein the processing unit performs a deep learning system according to the plurality of first sample images before and after the amplification Training a program to establish an image feature automatic identification algorithm, the image feature automatic identification algorithm comprising a first identification standard for the first image feature; and a detecting end, The image sensor is connected to the server, and the image feature automatic recognition algorithm is received by the server. The detecting end includes: an image capturing module, which is used to capture a to-be-detected object. Measure image; a processing module, which is connected to the image capturing module signal to obtain the image to be tested from the image capturing module, and the processing module performs the image feature automatic identification algorithm Determining, according to the first identification criterion, whether the image to be tested has the first image feature, wherein the processing module of the detecting end determines a plurality of the to-be-processed When the measured images do not conform to the standard image, but each has a second image feature, each of the plurality of image to be tested having the second image feature is recorded as a plurality of second sample images, respectively. And providing a plurality of the second sample images to the server; wherein the processing unit of the server establishes a second image feature category group in the database, and the plurality of the first The two sample images are stored in the second image feature category group. The system of claim 7, wherein the image processing program comprises one or a combination of two or more of an image shape adjustment program, an image contrast adjustment program, an image grayscale adjustment program, and an image color temperature adjustment program. The system of claim 7, wherein the processing unit further performs the program of: at least one sample image of the plurality of the first sample images before performing a training program of a deep learning system Selecting a first verification image, and executing a training program of the deep learning system with the remaining plurality of first sample images; and after the depth learning system establishes the image feature automatic recognition algorithm, according to The first identification criterion determines whether the first verification image has the first image feature to confirm the correctness of the first identification criterion. The system of claim 9, wherein the number of the first verification images is a plurality, and the number of the first verification images accounts for 3% to 50% of the number of the plurality of first sample images, and the processing unit further performs the following procedure: setting a first correct rate threshold And performing the image feature automatic identification algorithm, and determining, according to the first identification criterion, whether the standard image and the plurality of first verification images have the first image feature, according to the determination result Recording, by the first identification criterion, a first identification accuracy rate of the first image feature; wherein, when the first recognition accuracy rate is lower than the first correct rate threshold, performing the following steps: The first sample image is again subjected to the image processing program to further generate the amplified plurality of first sample images; and from the plurality of first sample images after the re-amplification, the third sample is again selected. % to 50% is the first verification image, and the remaining plurality of first sample images are again supplied to the depth learning system to correct the first identification criterion, and the corrected first An identification criterion again judges the standard image to And a plurality of the first verification images having the first image feature, and obtaining the first recognition accuracy rate again; wherein, the first recognition accuracy rate reaches the first correct rate threshold The image feature automatic recognition algorithm is provided to the detecting end. The system of claim 7, wherein the processing unit further performs the following program: reading a plurality of the second sample images in the database, and respectively performing the image processing programs, and respectively And superimposing the image processing results generated by the image processing program with the standard image to generate a plurality of amplified second sample images; and according to the plurality of pre-amplification and post-amplification The product image executes a training program of the depth learning system to be built in the image feature automatic recognition algorithm Establishing a second identification criterion for the second image feature. The system of claim 11, wherein the processing module determines the image to be tested according to the first identification criterion when the image feature automatic recognition algorithm analyzes the image to be tested. Whether the image has the first image feature, and determining whether the image to be tested has the second image feature according to the second identification criterion. The system of claim 7, further comprising: a sample image supply end coupled to the servo end signal to provide a plurality of the first sample images to the servo end; wherein The processing unit of the server further performs the following procedure: selecting 3% to 50% of the plurality of the first sample images as the first pre-examination image, and the first sample image And remaining a plurality of first sample images in the image are provided to the depth learning system to establish a first pre-test criterion including the first image feature; setting a first pre-check threshold; and Determining, by the first pre-test standard, whether the standard image and the plurality of first verification images have the first image feature, and recording the first pre-test standard to the first image according to the determination result a first pre-test correct rate of the feature; wherein, when the first pre-test correct rate is lower than the first pre-test threshold, more different first images are received from the sample image supply end The first sample image of the feature; wherein, in Said first pre-inspection when the correct rate of the first pre-inspection threshold, executing the image processing program. An image feature automatic identification method includes the following steps: Performing an image processing procedure on the plurality of first sample images respectively having different first image features, and separately performing image processing results generated by the image processing program and a standard not having the first image features The images are superimposed to generate a plurality of amplified first sample images; and the plurality of first sample images before and after the amplification are provided to a deep learning system to establish an automatic image feature recognition The algorithm, the image feature automatic identification algorithm includes a first identification criterion for the first image feature; capturing a to-be-tested image of the object to be detected, and automatically determining the algorithm pair by the image feature Performing analysis on the image to be tested, and determining, according to the first identification criterion, whether the image to be tested has the first image feature; and determining that the plurality of images to be tested do not meet the standard image For example, when all have a second image feature, a plurality of the images to be tested are respectively recorded as a plurality of second sample images. The method of claim 14, wherein the image processing program comprises one or a combination of two or more of an image shape adjustment program, an image contrast adjustment program, an image grayscale adjustment program, and an image color temperature adjustment program. The method of claim 14, further comprising: at least one sample image of the plurality of the first sample images before providing the plurality of first sample images to the depth learning system Selecting a first verification image and providing the remaining plurality of first sample images in the first sample image to the depth learning system; and establishing the image feature in the depth learning system After the automatic recognition algorithm is performed, determining whether the first verification image has the first image feature according to the first identification criterion to confirm the correctness of the first identification criterion. The method of claim 16, wherein the number of the first verification images is plural, and the number of the first verification images accounts for 3 of the plurality of first sample images. % to 50%, the method further comprising: setting a first correct rate threshold; and performing the image feature automatic identification algorithm, and determining the standard image and the plurality of locations according to the first identification criterion Whether the first verification image has the first image feature, and according to the determination result, recording a first recognition accuracy rate of the first image standard to the first image feature; wherein, the first identification is correct When the rate is lower than the first correct rate threshold, performing the following steps: performing the image processing procedure on the plurality of first sample images to further generate the amplified plurality of first sample images; From the plurality of first sample images after re-amplification, 3% to 50% are again selected as the first verification image, and the remaining plurality of first sample images are again supplied to the deep learning System to correct the first identification criteria and to correct The first identification criterion again determines whether the standard image and the plurality of the first verification images have the first image feature, and obtains the first recognition accuracy rate again; wherein, in the When the first identification accuracy rate reaches the first correct rate threshold, determining whether the image to be detected extracted from the object to be detected has the first image feature according to the first identification criterion. The method of claim 14, further comprising: performing the image processing procedure on the plurality of second sample images respectively having different second image features, and separately processing the image generated by the image processing program The result is superimposed with the standard image without the second image feature to generate a plurality of amplified second sample images; and the plurality of second sample images before and after the amplification are provided to The depth a system for establishing a second identification criterion for the second image feature in the image feature automatic recognition algorithm; and capturing the image to be detected of the object to be detected, the image feature The automatic identification algorithm analyzes the image to be tested, and determines whether the image to be tested has the first image feature according to the first identification criterion; wherein, the second identification criterion is determined according to the second identification criterion Determining whether the image to be tested has the second image feature. The method of claim 14, wherein before performing the image processing program, a pre-verification program is executed, the pre-verification program comprising: 3% to a plurality of the first sample images 50% is selected as the first pre-examination image, and the remaining plurality of first sample images in the first sample image are provided to the deep learning system to establish including the first a first pre-verification criterion of the image feature; setting a first pre-test threshold value; and determining, according to the first pre-verification criterion, whether the standard image and the plurality of the first pre-test images have the Determining, by the first image feature, a first pre-test correct rate of the first image feature according to the determination result; wherein, the first pre-test correct rate is lower than the first pre-test When the threshold value is verified, the number of the first sample images having different first image features is increased; and wherein, when the first pre-test correct rate reaches the first pre-test threshold, the execution The image processing program.