CN111986090B - Novel image data expansion method, system, terminal and storage medium - Google Patents
Novel image data expansion method, system, terminal and storage medium Download PDFInfo
- Publication number
- CN111986090B CN111986090B CN202010895511.XA CN202010895511A CN111986090B CN 111986090 B CN111986090 B CN 111986090B CN 202010895511 A CN202010895511 A CN 202010895511A CN 111986090 B CN111986090 B CN 111986090B
- Authority
- CN
- China
- Prior art keywords
- image
- spliced
- original
- training set
- label
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformations in the plane of the image
- G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
- G06T3/4038—Image mosaicing, e.g. composing plane images from plane sub-images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The invention belongs to the technical field of machine vision, and discloses a novel image data expansion method, which uses a training set organization method and a training set image stitching method to expand an original image training set to obtain a stitching image training set and stitching labels corresponding to each stitching image; and training the image classifier by using the spliced image training set and the spliced label. And splicing the original image test set by using a test set organization method and a test set image splicing method to obtain a spliced image test set. The invention uses different organization methods from the training set in the test set, does not increase the sample size of the spliced image test set, and improves the prediction efficiency of the image classifier. The image classifier trained by the spliced image training set can only output probability distribution on the spliced label, and the label mapping function maps the output probability distribution on the original label, so that the prediction of the test original image is completed.
Description
Technical Field
The invention belongs to the technical field of machine vision, and particularly relates to a novel image data expansion method, system, terminal and storage medium.
Background
Currently, deep convolutional neural networks have achieved tremendous success in image recognition. Unlike conventional feature engineering, deep convolutional neural networks reduce the dimension of an image to a one-dimensional vector through the convolutional layer and the pooling layer, and thus no related expertise is required to design a series of features. Deep convolutional neural networks require a large number of parameters to fit the mapping between inputs and outputs. Because of the large number of parameters, a large volume of annotated image data is required to train and optimize these parameters, which would otherwise be overfitted. In some fields, such as drug discovery, medical imaging, image data is scarce, and labeling of the data must be done by professionals, so that large labeled data sets are difficult to collect in these fields. Data expansion is a technique that can increase the sample size of a data set. In deep neural network applications, data expansion is often used to increase the sample size of the training set, which can well prevent overfitting.
The data expansion techniques commonly used at present include: flipping, clipping, padding, rotating, adding noise, gaussian blurring, affine transformation, generative antagonism network, etc. The existing data expansion techniques have the following problems and disadvantages:
1) Introducing false information and destroying real information
Existing data augmentation techniques generate new images by modifying part of the information of the real image. The manner of modification can be divided into two types: the relative positional relationship between the pixel values is modified and changed. These modifications make the newly generated image different from the real image, but at the same time introduce false information, resulting in many images being generated that are not present in the real world. For example, it is difficult to find a face in a crowd that matches a face image after affine transformation. During training, this false information can mislead the learning of the image classifier.
2) Part of the technical operation process is complex
The operation process of the existing partial data expansion technology is very complex. For example, in gaussian blur, the gaussian kernel needs to be slid over the entire image, and each sliding is followed by a multiplication summation with the area covered by the gaussian kernel. On a 100×100 image, if gaussian blur processing is performed, 10000 times of multiplication and summation are required. There is also the most popular generation of antagonistic networking technology in which deep learning is used, and the training process requires a great deal of computation and time. These data expansion techniques with complex operations all have an adverse effect on the efficiency of the system.
3) Destroying some of the potential relevance of image information
In some analysis of medical images, it is necessary to find areas from the image that are relevant to certain indices. If the found region is added with false information by existing data augmentation techniques, the confidence of this finding is questioned. Since we cannot determine how much of this discovery is affected by spurious information.
The difficulty of solving the problems and the defects is as follows:
under the thinking of the existing data expansion technology, a new image which has many similarities with the real image but at the same time has differences is expected to be generated. Therefore, the current data expansion technology is based on the real image, and ensures that the new image and the real image have the same point; and false information is introduced into the real image, so that the difference between the new image and the real image is ensured. Therefore, a brand new data expansion idea needs to be provided to thoroughly avoid the introduction of false information. And the data expansion technology generated under the thought cannot have an excessively complex operation process, so that adverse effects on efficiency are prevented.
The meaning of solving the problems and the defects is as follows:
False information is not introduced into the image any more, so that the image classifier can be trained most correctly and is not misled by the false information. On the other hand, if there is no complex operation in the data expansion technique, efficiency can be ensured. Finally, the potential relation between the image and certain indexes can be reserved.
Based on the defects, the invention provides a new image data expansion method. In the training stage, the method obtains a new spliced image by splicing a plurality of original images belonging to the same category in the original image training set, and configures corresponding spliced labels for the spliced image. And in the test stage, repeatedly splicing an original image in the original image test set to obtain a spliced image, and mapping the probability distribution of the spliced image on the spliced label to the original label by using a label mapping function. The method does not change the information of the original images, and the information in the spliced images is all from a plurality of original images.
Disclosure of Invention
Aiming at the problems existing in the prior art, the invention provides a novel image data expansion method.
The invention is realized in such a way that a new image data expansion method comprises the following three steps:
Step 1: and expanding the original image training set by using a training set organization method and a training set image stitching method to obtain a stitched image training set and a stitching label thereof.
1) Training set organization method
Randomly selecting a plurality of original images belonging to the same category in an original image training set to splice together to obtain a spliced image, and configuring a corresponding spliced label for each spliced image; all the spliced images together form a spliced image training set.
Assuming that m represents the number of categories in the training set of original images, N i (i=1, …, m) represents the number of original images belonging to category i in the training set of original images, L i (i=1, …, m) represents the label of category i, k represents the number of original images required to obtain one stitched image,Representing a splice of labels. The sample size N of the original image training set can be expressed as:
Splice tags can be expressed as:
Wherein the method comprises the steps of And (5) representing a spliced label obtained by splicing k images of L i.
Belonging to splice labelsThe number of stitched images of (a) is:
the sample size N s of the stitched image training set is:
the sample size of the visible training set is extended from N to The magnitude of the expansion is controlled by the parameter k, the larger the k value the larger the magnitude of the expansion.
2) Training set image stitching method
And connecting a plurality of original images end to end in the width direction, and adding a spacing area at the connecting position, wherein all pixel values in the spacing area are 0.
Assuming that the width of the original image is w, the height is h, the number of original images required for one stitched image is k, and the width of the interval region is g, the width of the stitched image is kw+ (k-1) g, and the height is still h.
Step 2: and splicing the original image test set by using a test set organization method and an image splicing method to obtain a spliced image test set.
1) Test set organization method
The purpose of data expansion is to increase the sample size of the training set, not the test set, so that in the test set, there is no need to obtain a stitched image by stitching multiple different original images to increase the sample size. The images in the spliced image test set and the images in the spliced image training set only need to be spliced repeatedly so that the images in the spliced image test set and the images in the spliced image training set have the same width and height.
And repeating the original image k times in the original image test set, and then splicing the original images together to obtain corresponding spliced images. And splicing the original images to obtain a spliced image. All the spliced images form a spliced image test set. The sample size of the spliced image test set is consistent with that of the original image test set.
2) Test set image stitching method
Repeating the same original image k times, connecting the end to end, and adding a spacing area at the connecting part, wherein all pixel values in the spacing area are 0.
The width and the height of the obtained spliced image are consistent with those of the spliced image in the spliced image training set.
Step 3: the probability distribution of the stitched image test set on the stitched labels is mapped onto the original labels using a label mapping function.
In the task of image recognition, what is ultimately needed is the probability distribution of all the original images in the test set on the original label. The image classifier outputs probability distribution of the spliced image on the spliced label, so that the probability distribution of the spliced image test set on the spliced label needs to be mapped onto the original label by using a label mapping function.
It is assumed that the number of the sub-blocks,Representing stitched images in a stitched image test setP j(Li), (j=1, …, k; i=1, …, m) represents the probability distribution of the j-th original image in the stitched image on the original label L i.K P j(Li) and thus can be obtained
Since the k original images are all identical, it is possible to obtain:
Thereby completing the mapping of the probability distribution from the spliced label to the original label.
Another object of the present invention is to provide a new image data expansion system, including:
The training module expands the original image training set by using a training set organization method and a training set image stitching method to obtain a stitching image training set and stitching labels corresponding to each stitching image; training the image classifier by using the spliced image training set and the spliced label;
And the testing module is used for splicing the original image testing set by using a testing set organization method and a testing set image splicing method to obtain a spliced image testing set. Inputting the spliced image test set into a trained image classifier to obtain probability distribution of each test spliced image on the spliced label;
And the mapping module is used for mapping the probability distribution of the spliced labels to the probability distribution of the original labels by using a label mapping function.
Another object of the present invention is to provide a drug discovery, medical image field image processing terminal to which the new image data expansion method is applied, the terminal being mounted with an image classifier, the image classifier executing a computer program which, when executed by the image classifier, causes the following steps to be performed:
In the training stage, an original image training set is expanded by using a training set organization method and a training set image stitching method to obtain a stitching image training set and stitching labels corresponding to each stitching image; training the image classifier by using the spliced image training set and the spliced label;
In the test stage, splicing the original image test set by using a test set organization method and a test set image splicing method to obtain a spliced image test set; inputting the spliced image test set into a trained image classifier to obtain probability distribution of each test spliced image on the spliced labels, and then mapping the probability distribution of the spliced labels onto the probability distribution of the original labels by using a label mapping function.
It is a further object of the present invention to provide a computer device comprising a memory and a processor, the memory storing a computer program which, when executed by the processor, causes the processor to perform the steps of:
In the training stage, an original image training set is expanded by using a training set organization method and a training set image stitching method to obtain a stitching image training set and stitching labels corresponding to each stitching image; training the image classifier by using the spliced image training set and the spliced label;
In the test stage, splicing the original image test set by using a test set organization method and a test set image splicing method to obtain a spliced image test set; inputting the spliced image test set into a trained image classifier to obtain probability distribution of each test spliced image on the spliced labels, and then mapping the probability distribution of the spliced labels onto the probability distribution of the original labels by using a label mapping function.
Another object of the present invention is to provide a computer readable storage medium storing a computer program which, when executed by a processor, causes the processor to perform the steps of:
In the training stage, an original image training set is expanded by using a training set organization method and a training set image stitching method to obtain a stitching image training set and stitching labels corresponding to each stitching image; training the image classifier by using the spliced image training set and the spliced label;
In the test stage, splicing the original image test set by using a test set organization method and a test set image splicing method to obtain a spliced image test set; inputting the spliced image test set into a trained image classifier to obtain probability distribution of each test spliced image on the spliced labels, and then mapping the probability distribution of the spliced labels onto the probability distribution of the original labels by using a label mapping function.
By combining all the technical schemes, the invention has the advantages and positive effects that:
the expansion of the image data is achieved by stitching the original images into a new stitched image. The method is applied to image recognition and can be matched with various image classifiers for use. And in the training stage, the original image training set is expanded by using a training set organization method and a training set image stitching method to obtain a stitching image training set and stitching labels corresponding to each stitching image. And training the image classifier by using the spliced image training set and the spliced label. And in the test stage, splicing the original image test set by using a test set organization method and a test set image splicing method to obtain a spliced image test set. Inputting the spliced image test set into a trained image classifier to obtain probability distribution of each test spliced image on the spliced labels, and then mapping the probability distribution of the spliced labels onto the probability distribution of the original labels by using a label mapping function.
Advantages of the present invention compared to the prior art further include:
1) Existing data augmentation methods all generate new images by modifying the information of the original image. According to the invention, a new spliced image is obtained by splicing a plurality of original images, and all the original images are repeatedly arranged to obtain all the spliced images.
2) And in the test set, repeating the splicing of the original images k times to obtain a spliced image.
3) The stitching method of images does not stitch the original images together in the width direction directly, but adds a space region between the original images.
4) The probability distribution on the splice label is mapped onto the original label using a label mapping function.
5) The existing data expansion method can destroy the information of the original image and introduce false information into the new image, and all information of the newly generated spliced image comes from the original image without introducing any false information.
5) In the test set, a different organization method from the training set is used, the sample size of the spliced image test set is not increased, and the prediction efficiency of the image classifier is improved.
6) The addition of the interval region eliminates the mutual interference of two adjacent original image information in the spliced image.
7) The image classifier trained by the spliced image training set can only output probability distribution on the spliced label, and the label mapping function maps the output probability distribution on the original label, so that the prediction of the test original image is completed.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings that are needed in the embodiments of the present application will be briefly described below, and it is obvious that the drawings described below are only some embodiments of the present application, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
Fig. 1 is a flowchart of a new image data expansion method according to an embodiment of the present invention.
Fig. 2 is a schematic diagram of a new image data expansion method according to an embodiment of the present invention.
Fig. 3 is a schematic diagram of a splicing method according to an embodiment of the present invention;
FIG. 4 is a schematic diagram of an organization method and an image stitching method in an embodiment of the present invention.
FIG. 5 is a schematic diagram of label mapping in an embodiment of the invention.
Detailed Description
The present invention will be described in further detail with reference to the following examples in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.
In view of the problems existing in the prior art, the present invention provides a new image data expansion method, system, terminal and storage medium, and the present invention is described in detail below with reference to the accompanying drawings.
As shown in fig. 1, an embodiment of the present invention provides a new image data expansion method, including:
S101: and expanding the original image training data set by using a training set organization method and a training set image stitching method to obtain a stitched image training set and a stitching label thereof.
S102: and splicing the original image test set by using a test set organization method and a test set image splicing method to obtain a spliced image test set.
S103: the probability distribution of the stitched image test set on the stitched labels is mapped onto the original labels using a label mapping function.
In step S101, the training data set of the original image is expanded by using a training set organization method and a training set image stitching method, so as to obtain a stitched image training set and a stitched label thereof.
1) Training set organization method
Randomly selecting a plurality of original images belonging to the same category in an original image training set to splice together to obtain a spliced image, and configuring a corresponding spliced label for each spliced image; all the spliced images together form a spliced image training set.
Assuming that m represents the number of categories in the training set of original images, N i (i=1, …, m) represents the number of original images belonging to category i in the training set of original images, L i (i=1, …, m) represents the label of category i, k represents the number of original images required to obtain one stitched image,Representing a splice of labels. The sample size N of the original image training set can be expressed as:
Splice tags can be expressed as:
Wherein the method comprises the steps of And (5) representing a spliced label obtained by splicing k images of L i.
Belonging to splice labelsThe number of stitched images of (a) is:
the sample size N s of the stitched image training set is:
the sample size of the visible training set is extended from N to The magnitude of the expansion is controlled by the parameter k, the larger the k value the larger the magnitude of the expansion.
2) Training set image stitching method
And connecting a plurality of original images end to end in the width direction, and adding a spacing area at the connecting position, wherein all pixel values in the spacing area are 0.
Assuming that the width of the original image is w, the height is h, the number of original images required for one stitched image is k, and the width of the interval region is g, the width of the stitched image is kw+ (k-1) g, and the height is still h.
In step S102, the original image test set is stitched by using a test set organization method and a test set image stitching method, so as to obtain a stitched image test set.
1) Test set organization method
The purpose of data expansion is to increase the sample size of the training set, not the test set, so that in the test set, there is no need to obtain a stitched image by stitching multiple different original images to increase the sample size. The images in the spliced image test set and the images in the spliced image training set only need to be spliced repeatedly so that the images in the spliced image test set and the images in the spliced image training set have the same width and height.
And repeating the original image k times in the original image test set, and then splicing the original images together to obtain corresponding spliced images. And splicing the original images to obtain a spliced image. All the spliced images form a spliced image test set. The sample size of the spliced image test set is consistent with that of the original image test set.
2) Test set image stitching method
Repeating the same original image k times, connecting the end to end, and adding a spacing area at the connecting part, wherein all pixel values in the spacing area are 0.
The width and the height of the obtained spliced image are consistent with those of the spliced image in the spliced image training set.
In step S103, the probability distribution of the stitched image test set on the stitched label is mapped onto the original label using a label mapping function.
In the task of image recognition, what is ultimately needed is the probability distribution of all the original images in the test set on the original label. The image classifier outputs probability distribution of the spliced image on the spliced label, so that the probability distribution of the spliced image test set on the spliced label needs to be mapped onto the original label by using a label mapping function.
It is assumed that the number of the sub-blocks,Representing stitched images in a stitched image test setP j(Li), (j=1, …, k; i=1, …, m) represents the probability distribution of the j-th original image in the stitched image on the original label L i.K P j(Li) and thus can be obtained
Since the k original images are all identical, it is possible to obtain:
Thereby completing the mapping of the probability distribution from the spliced label to the original label.
The present invention provides a new image data expansion system, the new image data expansion system comprising:
The training module expands the original image training set by using a training set organization method and a training set image stitching method to obtain a stitching image training set and stitching labels corresponding to each stitching image; training the image classifier by using the spliced image training set and the spliced label;
And the testing module is used for splicing the original image testing set by using a testing set organization method and a testing set image splicing method to obtain a spliced image testing set. Inputting the spliced image test set into a trained image classifier to obtain probability distribution of each test spliced image on the spliced label;
And the mapping module is used for mapping the probability distribution of the spliced labels to the probability distribution of the original labels by using a label mapping function.
The invention is further described below in connection with specific embodiments.
Example 1
The invention aims to expand the sample size of a training set by splicing a plurality of original images to obtain a new spliced image.
As shown in fig. 4, is a task of recognizing handwritten numbers in images, examples of which are from a common dataset MNIST. The numbers to be identified are "1" and "2", respectively, and the numbers in the image are the labels of the image. The original image training set has 4 images, two images of digital "1" and two images of digital "2". The original image test set has two images, the numbers in the images are unknown. The images in the original image training set and the original image testing set have the width of 28 pixels, the height of 28 pixels and the channel number of 1. The goal of this task is to train an image classifier through the training set to identify numbers in the images in the test set.
The sample size of the original image training set is only 4, and the original image training set is expanded by using the method of the invention. In the original image training set, 2 images are randomly selected and spliced according to the image label '1', so that a new spliced image is obtained. The number of original images with the label of "1" is 2, and the number of repeated arrangement is 4, so that the spliced image obtained after expansion is also 4. The splice label of the spliced image is "11". For the image label "2", 4 spliced images can be obtained, and the spliced label is "22". In the training set of stitched images, there are 8 stitched images, and the stitched labels are "11" and "22". The invention thus extends the sample size of the training set from 4 to 8.
Similarly, if 3 original images are spliced to obtain a spliced image, the sample size of the spliced image training set is 16; if 4 original images are spliced to obtain a spliced image, the sample size of the spliced image training set is 32; … …. Therefore, the sample size of the training set of the spliced image can be controlled by setting the number of the original images in one spliced image.
In the original image test set, in order to make the image size consistent with the image size in the spliced image training set, the images of the original image test set also need to be spliced. The sample size of the original image test set is 2, and the sample size does not influence the training of the image classifier, so that the original image test set does not need to be expanded. The corresponding spliced image can be obtained by only repeating the same original image twice. The sample size of the obtained spliced image test set is still 2, and compared with the original image test set, the sample size of the spliced image test set is unchanged.
After stitching, the size of the image may change. In this example, the width of the space region is 3, and thus the width of the stitched image is 28×2+3=59, the height is also 28×2+3=59, and the number of channels is 1. The size of the stitched image is also controlled by the number of original images in the stitched image.
After the image classifier is trained by the spliced image training set, the spliced images in the spliced image test set are predicted, and the prediction result is probability distribution on spliced labels 11 and 22. The aim of this task is to obtain a probability distribution of the original image on the original label. The probability distribution is mapped from "11", "22" to "1", "2" using a label mapping function.
In this embodiment, probability distributions of two stitched images in the stitched image test set on "11" and "22" are [0.9,0.1] and [0.2,0.8], respectively. The probability distribution of the stitched image is the joint probability of the probability distribution of the original images, so the probability distribution of the two original images on the "1", "2" is [0.949,0.316], [0.447,0.894]. The image classifier can be known to predict the number of 1 in the first original image and the number of 2 in the original image, so that the prediction of the original image in the test set is completed.
Fig. 3 is a schematic diagram of a splicing method according to an embodiment of the present invention; in fig. 3, n represents the sample size of the data set; in this example, 2 original images are stitched to obtain a stitched image.
FIG. 4 is a schematic diagram of an organization method and an image stitching method in an embodiment of the present invention. In fig. 4, n represents the sample size of the data set; in this example, 2 original images are stitched to obtain a stitched image.Representing interval intervals.
FIG. 5 is a schematic diagram of label mapping in an embodiment of the invention.
Example two
In this embodiment, we also compare the effect of the data expansion method of the present invention with other data expansion methods. The image data is from the MNIST dataset. The original image training set has a total of 5 categories of images, 5 images per category, and a total of 25 images. Three different data expansion modes are used for the original image training set: 1) The 25 original images are directly used for training the image classifier without expansion; 2) Expanding the training set by using the existing common data expansion method, expanding each original image to obtain 10 new images, and finally obtaining an expanded training set with a sample size of 275; 3) The training set is expanded by using the data expansion method in the invention, wherein the number k of the original images in the spliced image is 2,3,4 and 5 respectively, and the sample size of the spliced image training set correspondingly obtained is (5 multiplied by 5 2),(5×53),(5×54),(5×55) respectively. Then training a randomly initialized convolutional neural network by using the training sets, and testing the trained convolutional neural network by using the testing set. The test results are as follows.
Table 1 experimental results
| Method of | Classification accuracy |
| Data-free augmentation | 81.62％ |
| Existing data expansion method | 83.67％ |
| Data expansion method of the present invention (k=2) | 81.84％ |
| Data expansion method of the invention (k=3) | 86.59％ |
| Data expansion method of the invention (k=4) | 89.15％ |
| Data expansion method of the invention (k=5) | 90.81％ |
It can be seen that the classification accuracy of the data expansion method of the present invention is higher than that of the existing data expansion method.
In the description of the present invention, unless otherwise indicated, the meaning of "a plurality" is two or more; the terms "upper," "lower," "left," "right," "inner," "outer," "front," "rear," "head," "tail," and the like are used as an orientation or positional relationship based on that shown in the drawings, merely to facilitate description of the invention and to simplify the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore should not be construed as limiting the invention. Furthermore, the terms "first," "second," "third," and the like are used for descriptive purposes only and are not to be construed as indicating or implying relative importance.
It should be noted that the embodiments of the present invention can be realized in hardware, software, or a combination of software and hardware. The hardware portion may be implemented using dedicated logic; the software portions may be stored in a memory and executed by a suitable instruction execution system, such as a microprocessor or special purpose design hardware. Those of ordinary skill in the art will appreciate that the apparatus and methods described above may be implemented using computer executable instructions and/or embodied in processor control code, such as provided on a carrier medium such as a magnetic disk, CD or DVD-ROM, a programmable memory such as read only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The device of the present invention and its modules may be implemented by hardware circuitry, such as very large scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., as well as software executed by various types of processors, or by a combination of the above hardware circuitry and software, such as firmware.
The foregoing is merely illustrative of specific embodiments of the present invention, and the scope of the invention is not limited thereto, but any modifications, equivalents, improvements and alternatives falling within the spirit and principles of the present invention will be apparent to those skilled in the art within the scope of the present invention.
Claims (6)
1. A new image data expansion method, characterized in that the new image data expansion method comprises:
Expanding an original image training set by using a training set organization method and a training set image stitching method to obtain a stitched image training set and stitching labels corresponding to each stitched image;
training the image classifier by using the spliced image training set and the spliced label;
splicing the original image test sets by using a test set organization method and a test set image splicing method to obtain spliced image test sets; inputting the spliced image test set into a trained image classifier to obtain probability distribution of each test spliced image on the spliced labels, and mapping the probability distribution of the spliced labels onto the probability distribution of the original labels by using a label mapping function;
the training set organization method comprises the following steps:
Randomly selecting a plurality of original images belonging to the same category in an original image training set to splice together to obtain a spliced image, and configuring a corresponding spliced label for each spliced image; all the spliced images together form a spliced image training set;
m represents the number of categories in the training set of original images, N i (i=1, …, m) represents the number of original images belonging to category i in the training set of original images, L i (i=1, …, m) represents the label of category i, k represents the number of original images required to obtain one stitched image, Representing the concatenation of labels; the sample size N of the original image training set is expressed as:
The splice tag is expressed as:
Wherein the method comprises the steps of Representing a spliced label obtained by splicing k L i images;
Belonging to splice labels The number of stitched images of (a) is:
the sample size N s of the stitched image training set is:
The sample size of the training set is expanded from N to N s; the amplitude of expansion is controlled by the parameter k;
the training set image stitching method comprises the following steps:
Connecting a plurality of original images end to end in the width direction, and adding a spacing area at the connecting position, wherein all pixel values in the spacing area are 0;
The width of the original image is w, the height is h, the number of the original images required by one spliced image is k, the width of the interval area is g, the width of the spliced image is kw+ (k-1) g, and the height is h;
the test set organization method comprises the following steps:
repeating an original image k times in an original image test set, and then splicing the original images together to obtain corresponding spliced images; splicing an original image to obtain a spliced image; all the spliced images form a spliced image test set; the sample size of the spliced image test set is consistent with that of the original image test set;
The test set image stitching method comprises the following steps:
Repeating the same original image for k times, connecting the end to end, and adding a spacing area at the connecting part, wherein all pixel values in the spacing area are 0; the width and the height of the obtained spliced image are consistent with those of the spliced image in the spliced image training set.
2. The new image data augmentation method of claim 1, wherein the method of mapping probability distributions of a stitched image test set on a stitched label onto an original label using a label mapping function comprises:
Representing stitched images in a stitched image test set P j(Li), (j=1, …, k; i=1, …, m) represents the probability distribution of the jth original image in the stitched image on the original label L i; k P j(Li) to obtain
The k original images are the same, and the following results are obtained:
3. A new image data expansion system implementing the new image data expansion method according to any one of claims 1 to 2, characterized in that the new image data expansion system comprises:
The training module expands the original image training set by using a training set organization method and a training set image stitching method to obtain a stitching image training set and stitching labels corresponding to each stitching image; training the image classifier by using the spliced image training set and the spliced label;
The testing module is used for splicing the original image testing set by using a testing set organization method and a testing set image splicing method to obtain a spliced image testing set; inputting the spliced image test set into a trained image classifier to obtain probability distribution of each test spliced image on the spliced label;
And the mapping module is used for mapping the probability distribution of the spliced labels to the probability distribution of the original labels by using a label mapping function.
4. A drug discovery, medical image domain image processing terminal applying the new image data expansion method according to any one of claims 1-2, characterized in that said terminal is equipped with an image classifier, said image classifier executing a computer program, which when executed by said image classifier causes the execution of the steps of:
In the training stage, an original image training set is expanded by using a training set organization method and a training set image stitching method to obtain a stitching image training set and stitching labels corresponding to each stitching image; training the image classifier by using the spliced image training set and the spliced label;
In the test stage, splicing the original image test set by using a test set organization method and a test set image splicing method to obtain a spliced image test set; inputting the spliced image test set into a trained image classifier to obtain probability distribution of each test spliced image on the spliced labels, and then mapping the probability distribution of the spliced labels onto the probability distribution of the original labels by using a label mapping function.
5. A computer device comprising a memory and a processor, the memory storing a computer program which, when executed by the processor, causes the processor to perform the steps of the new image data expansion method according to any of claims 1-2.
6. A computer readable storage medium storing a computer program which, when executed by a processor, causes the processor to perform the steps of the new image data expansion method according to any one of claims 1-2.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010895511.XA CN111986090B (en) | 2020-08-31 | 2020-08-31 | Novel image data expansion method, system, terminal and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010895511.XA CN111986090B (en) | 2020-08-31 | 2020-08-31 | Novel image data expansion method, system, terminal and storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111986090A CN111986090A (en) | 2020-11-24 |
| CN111986090B true CN111986090B (en) | 2024-07-05 |
Family
ID=73440495
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010895511.XA Active CN111986090B (en) | 2020-08-31 | 2020-08-31 | Novel image data expansion method, system, terminal and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111986090B (en) |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109710800A (en) * | 2018-11-08 | 2019-05-03 | 北京奇艺世纪科技有限公司 | Model generating method, video classification methods, device, terminal and storage medium |
| CN111126514A (en) * | 2020-03-30 | 2020-05-08 | 同盾控股有限公司 | Image multi-label classification method, device, equipment and medium |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11030486B2 (en) * | 2018-04-20 | 2021-06-08 | XNOR.ai, Inc. | Image classification through label progression |
-
2020
- 2020-08-31 CN CN202010895511.XA patent/CN111986090B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109710800A (en) * | 2018-11-08 | 2019-05-03 | 北京奇艺世纪科技有限公司 | Model generating method, video classification methods, device, terminal and storage medium |
| CN111126514A (en) * | 2020-03-30 | 2020-05-08 | 同盾控股有限公司 | Image multi-label classification method, device, equipment and medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111986090A (en) | 2020-11-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Chen et al. | Reverse attention-based residual network for salient object detection | |
| Dvornik et al. | On the importance of visual context for data augmentation in scene understanding | |
| Luo et al. | Moran: A multi-object rectified attention network for scene text recognition | |
| Mopuri et al. | CNN fixations: An unraveling approach to visualize the discriminative image regions | |
| Xiao et al. | Deep salient object detection with dense connections and distraction diagnosis | |
| Li et al. | Contrast-oriented deep neural networks for salient object detection | |
| US11640714B2 (en) | Video panoptic segmentation | |
| Xiao et al. | A weakly supervised semantic segmentation network by aggregating seed cues: the multi-object proposal generation perspective | |
| Chen et al. | Embedding attention and residual network for accurate salient object detection | |
| Xiong et al. | Diagnose like a pathologist: Transformer-enabled hierarchical attention-guided multiple instance learning for whole slide image classification | |
| Rahman et al. | A framework for fast automatic image cropping based on deep saliency map detection and gaussian filter | |
| CN111160140B (en) | Image detection method and device | |
| Jiang et al. | A CNN model for semantic person part segmentation with capacity optimization | |
| CN112861934A (en) | Image classification method and device of embedded terminal and embedded terminal | |
| Nemade et al. | Image segmentation using convolutional neural network for image annotation | |
| Huang et al. | ORDNet: Capturing omni-range dependencies for scene parsing | |
| Jiang et al. | Super diffusion for salient object detection | |
| CN115115825B (en) | Method, device, computer equipment and storage medium for detecting object in image | |
| Dakhia et al. | A hybrid-backward refinement model for salient object detection | |
| CN111986090B (en) | Novel image data expansion method, system, terminal and storage medium | |
| Zhang et al. | A componentwise approach to weakly supervised semantic segmentation using dual-feedback network | |
| CN113743434A (en) | Training method of target detection network, image augmentation method and device | |
| US20230162519A1 (en) | Memory-efficient feature descriptors for localization and classification of identity documents | |
| Shuhan et al. | Semantic image segmentation using region-based object detector | |
| Wang et al. | Polyp Segmentation via Semantic Enhanced Perceptual Network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |