CN110570393A - A method for detecting defects in the window area of mobile phone glass cover based on machine vision - Google Patents

Abstract

The invention discloses a method for detecting defects in the window area of a glass cover plate of a mobile phone based on machine vision, comprising the following steps: step 1, collecting an image of the mobile phone screen; step 2, performing a rough inspection on the screen area of the mobile phone; step 3, passing a threshold segmentation algorithm Extract the defects of the mobile phone screen image; Step 4, use the clustering algorithm to connect the scattered points in the dense point cluster area; Step 5, use the neural network classifier to classify the defects; Step 6, Extract the area, length and radius of the defect area , and compare according to the detection standard; Step 7, use the deep learning classifier to reclassify the defects; Step 8, count the information and quantity of various types of defects. The detection algorithm of the invention follows the principle of rough inspection first and fine inspection later, which can not only quickly and accurately extract defects such as pitting, scratches, dirt, wool and other defects in the screen area of the glass cover of various types of mobile phones, but also Adjust the detection accuracy according to the detection standards of different product window areas.

Description

A method for detecting defects in the window area of mobile phone glass cover based on machine vision

technical field

The invention relates to the field of visual inspection, in particular to a method for detecting defects in the window area of a glass cover plate of a mobile phone based on machine vision.

Background technique

In recent years, driven by technologies such as 5G and wireless charging, non-metal mobile phone covers have become mainstream. Among them, the glass cover has better mechanical properties and optical properties, and the cost is lower than that of ceramic materials, so it is favored by 3C product companies. However, in the process of actual manufacturing, transportation, etc., defects such as pitting, scratches, dirt, and edge collapse are inevitable. Timely defect detection in the production process can avoid process waste and monitor the production quality of products, thereby ensuring the production of high-quality products and saving production costs.

At present, many domestic cover glass manufacturers still use a large number of manual visual inspection methods. Compared with the manual visual inspection process, machine vision inspection technology has higher inspection efficiency, lower cost, and more stable inspection standards. At present, many domestic testing equipment manufacturers have successively developed glass detection algorithms, but there are still problems such as unstable detection process, high missed detection or over-detection rate.

Deep learning technology has been widely used in speech recognition, target tracking, security monitoring and other fields. In terms of machine vision, deep learning also has great potential in defect classification and localization.

SUMMARY OF THE INVENTION

In order to solve the above problems, the present invention provides a method for detecting defects in the window area of mobile phone glass cover plates based on machine vision. It can quickly and accurately extract defects such as pitting, scratches, dirt and other defects in the screen area of 2D or 2.5D mobile phone glass cover of various models.

The present invention is realized by at least one of the following technical solutions.

A method for detecting defects in the window area of a glass cover of a mobile phone based on machine vision, comprising the following steps:

Step 1. Collect mobile phone screen images;

Step 2. Perform a rough inspection on the screen area of the mobile phone;

Step 3: Extract the defects of the mobile phone screen image through a threshold segmentation algorithm;

Step 4, using the clustering algorithm to connect the scattered points of the dense point cluster area on the mobile phone screen window area;

Step 5: Extract the connected domain features of the defects, input the connected domain features into the neural network classifier, and use the neural network classifier to classify point-like, linear and planar defects;

Step 6: Extract the area, length and radius of the defect area for the point-like, linear-like defect and plane-like defect areas, and compare them according to the detection standard;

Step 7. Use the deep learning classifier to reclassify the classified defects, and classify four types of defects: scratches, floating hairs, light-colored dirt, and dark-colored dirt;

Step 8: Statistics of various types of defect information and quantity.

Further, the first step is to use a 16K line scan camera to perform image acquisition on the mobile phone screen, the image resolution is 40000×16384, and the detection accuracy is 0.005mm.

Further, before performing the rough inspection, select a picture of a defect-free good-quality mobile phone screen, frame the mobile phone screen area as the region of interest (ROI) of the standard template, and save the standard template file offline; the standard template includes shape matching. Template and standard area template, among which, the shape matching template is mainly used to match the mobile phone screen in the image to be detected, and obtain the position information of the mobile phone screen in the image to be detected; the standard area template is mainly used to extract the information from the image to be detected The mobile phone screen area is aligned to detect defects;

The rough inspection of the mobile phone screen area in step 2 includes the following steps:

(1) During online detection, the shape matching template is read, and the shape-based template matching algorithm is used to match and obtain the position information and angle information of the mobile phone screen in the image to be detected;

(2) Use the threshold segmentation algorithm to segment the window area of the mobile phone screen in the image to be detected, and create an affine transformation matrix by obtaining the obtained position and angle information; read the standard area template of the window area, and use the area affine transformation to read The standard area template taken is aligned with the window area to be detected;

The aligned standard region template and the to-be-detected window region are subjected to Boolean operation to detect whether there is a defect; the Boolean operation is the operation of intersection, union and complement of the regions.

Further, step 3 extracting the defects of the mobile phone screen image includes the following steps:

1), use median filtering for image preprocessing to reduce noise;

2), dividing the window area to be detected, that is, the screen area, into m×n areas;

3), use the Gaussian filter enhancement algorithm to enhance the image of each segmented area. The Gaussian filter enhancement algorithm is:

img_enhanced=round((orig-mean_gauss)*factor)+orig(1)

In formula (1), mean_gauss represents the Gaussian filtering of the image, orig represents the corresponding gray value of each image, factor is the enhancement factor, and img_enhanced is the gray value of the enhanced output image; the round() function is a rounding function;

4), extract the defects in the window area of the mobile phone screen from the enhanced picture through the threshold segmentation algorithm.

Further, the clustering algorithm described in step 4 is to connect only the dense point regions that meet the clustering conditions, and the regions that do not meet the clustering conditions are not clustered and connected;

The clustering conditions are:

a. A circular dense point area with a clustering radius of R;

b. In the circular area, the total number of areas where the area of the dense point area is greater than the area S of the minimum area of dense points is greater than the number of clusters N;

In a circular area with a radius of R, if the number of areas that satisfy the requirement that the area of the dense point area is greater than the set minimum area S of the dense point area is greater than N, the dense point area is considered to be a dense point cluster area; using morphological expansion, The erosion operation connects the scatter points of this region into a region.

Further, the neural network classifier is used to classify point defects, linear defects and planar defects;

The neural network classifier includes an input layer, an intermediate layer and a hidden layer; the number of neurons in the input layer is the same as the dimension of the input feature vector, and the number of neurons in the output layer is the same as the classification output category. The hidden layer abstracts the features of the input data into another dimension space to show its more abstract features, which can be better linearly divided. Hidden layers can improve the accuracy and expressiveness of the network.

The neural network classifier for defect classification is mainly divided into two parts: offline training and online detection;

Before classifying the defects, the connected domain features of the defects extracted in step 5 are used as the defect data set; the connected domain features saved in the data set are divided into three categories, the first category is the basic features of area, center, width and height; The second category is the circumscribed circle radius, roundness and compact shape features; the third category is the geometric moment features of the second and third order moments;

Offline training: Set training parameters, including optimization algorithm, learning rate, learning strategy, initialization strategy, offline training on defective datasets, forward propagation to calculate loss function, and back propagation to update weights. Until the loss function reaches the convergence state; save the trained neural network classifier offline;

Online detection: Extract the connected domain features of defects, read the trained neural network classifier, and input the extracted defect connected domain features into the classifier to obtain the defect classification results.

Further, the area of the defect area described in step 6 is obtained by counting the number of pixels in the connected domain, which is used for the comparison of the defect standard of the planar class; the length of the defect area is obtained by skeletonizing the defect area, which is used for the line The defect length standard comparison of the shape type; the radius of the defect area is the minimum circumcircle radius of the defect area, which is used for the comparison of the defect radius standard.

Further, the detection standard of step 6 is:

1) Point defects

D<0.1mm, ignore;

0.1mm≤D≤0.2mm, d≥10mm, if the number of defects within the radius of d is less than 2, it means that the mobile phone screen is a good product, and if the number of defects is greater than 2, it means that the mobile phone screen is a defective product;

D>0.2mm, NG, indicating that the mobile phone screen is a defective product;

Among them, D is the diameter of the minimum circumscribed circle of point defects, and d is the distance between points;

2) Defective linearity

W<0.02mm, the defect is ignored;

0.02mm≤W≤0.05mm, d≥5mm, L≤3, if the defect can be ignored within 2, it means that the mobile phone screen is a good product; less than 3 means that the mobile phone screen is a defective product;

W>0.05mm or L>3mm, NG, indicating that the mobile phone screen is a defective product;

Among them, W is the width of the linear defect, L is the length of the linear defect, and d is the distance between different linear defects;

3) Poor face shape

D>0.5mm or S>0.2mm ²

Among them, D is the diameter of the largest inscribed circle of the planar defect, and S is the area of the planar defect.

Further, the deep learning classifier is a convolutional neural network model, and the convolutional neural network model mainly includes an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer, wherein the input layer, the convolutional layer , the pooling layer, the fully connected layer and the output layer are connected in turn; the convolutional layer completes the feature extraction operation, and converts the defect features of the last pooling layer into a set of vectors as the input of the fully connected layer, and the defect features of the fully connected layer Input to the output layer to complete the task of classification;

Defect classification by deep learning classifier is mainly divided into two parts: offline training and online detection;

Before classifying defects, create data sets of different types of defects, and use mirroring and rotation to enhance defect pictures in the data set;

The VGG16 convolutional neural network model of the Visual Geometry Group of Oxford University was used to set the model structure parameters and training parameters;

During offline training, use the pre-training model to perform migration training on the dataset; after training, save the deep learning classifier offline;

During online detection, the image corresponding to the defect area to be classified is intercepted, and the image is used as the input of the deep learning classifier to classify the defects.

Further, a parameter configuration file is made for the algorithm parameters and detection standards. The parameters can be flexibly adjusted according to the detection requirements, and it is suitable for the detection of the window area of different types of mobile phone glass cover products.

The algorithm parameters are mainly key parameters in various detection algorithms, such as the threshold in the threshold segmentation algorithm, the number of blocks and the cluster radius in the clustering algorithm.

The detection standard parameters are mainly the standard parameters set according to the detection requirements, such as the radius, length, width, and number of defects involved in the point, line, and plane detection standards.

Further, the present invention follows the principle of "rough inspection first, then fine inspection". Through the rough inspection process, the serious defects in the window area can be detected first, and through the fine inspection process, the classification of point-like linear and planar defects can be realized. For indistinguishable scratches, floating filaments, light-colored dirt, and dark-colored dirt, a deep learning classifier is used to reclassify defects.

The beneficial effects of the invention are as follows: by combining offline creation of templates with online detection, the detection algorithm has high versatility and can be applied to the detection of glass cover plates of different types of mobile phones; the principle of “rough inspection first, then fine inspection” is followed. , the algorithm has high detection efficiency; the combination of traditional detection algorithm and deep learning technology has a good classification effect on scratches, wool, dark stains, light stains and other difficult to distinguish defects; algorithm parameters and detection standard parameters A parameter configuration file is made, and the detection accuracy is flexible and adjustable.

Description of drawings

1 is a schematic flowchart of the detection process of a method for detecting defects in the window area of a glass cover of a mobile phone based on machine vision of the present invention;

Fig. 2 is the schematic diagram of window division block of the present invention;

3 is a schematic diagram of clustering conditions of the present invention;

4 is a schematic diagram of the clustering effect of the present invention;

Fig. 5 is the neural network classifier classification schematic diagram of the present invention;

Fig. 6 is the neural network classifier structure diagram of the present invention;

Fig. 7a is the defect schematic diagram of the scratch of the present invention;

Fig. 7b is the defect schematic diagram of the floating filament of the present invention;

Fig. 7c is the defect schematic diagram of the light-colored dirt of the present invention;

Fig. 7d is the defect schematic diagram of dark stain of the present invention;

8 is a schematic structural diagram of a convolutional neural network model of the present invention;

Fig. 9 is the schematic diagram of detection result of the present invention;

Figure 10 is a schematic diagram of the detection results of the present invention.

Detailed ways

The present invention will be further described in detail below in conjunction with specific embodiments.

The present invention will be described in further detail below with reference to the embodiments and the accompanying drawings, but the embodiments of the present invention are not limited thereto.

As shown in Figure 1, a method for detecting defects in the window area of the glass cover of a mobile phone based on machine vision mainly includes two parts: rough inspection and fine inspection. The rough inspection process includes four key steps: standard template making, template matching, affine transformation and region comparison. If there are serious defects, the next step is fine inspection. The fine inspection part mainly realizes the extraction and classification of defects. The region of interest (ROI) of the screen area of the image to be detected can be obtained according to the position and angle information matched in the rough inspection process. The defect extraction process includes four parts: image preprocessing, image segmentation, image enhancement and threshold segmentation. Before defect classification, the extracted defect areas are clustered; the defects are divided into point, line and plane defects by neural network classifier, and compared with the corresponding point, line and plane detection standards. right. Deep learning classifiers are used to reclassify scratches, floating hairs, light-colored dirt, and dark-colored dirt that are indistinguishable. Finally, statistical information such as defect area, position, length, width minimum circumscribed circle, maximum inscribed circle and so on.

A method for detecting defects in the window area of a glass cover of a mobile phone based on machine vision, comprising the following steps:

Step 1. Use a 16K line scan camera to collect the image of the mobile phone screen, the image resolution is 40000 × 16384, and the detection accuracy can reach 0.005mm.

Step 2. Perform a rough inspection on the screen area of the mobile phone;

Before the rough inspection, select a picture of a non-defective good mobile phone screen, and frame the mobile phone screen area as the region of interest (ROI) of the standard template. The standard template is mainly divided into two parts: a shape matching template and a standard area template. Among them, the shape matching template is mainly used to match the mobile phone screen in the image to be detected, and the position information of the mobile phone screen in the image to be detected is obtained; the standard area template is mainly used to extract the information from the image to be detected. The mobile phone screen area is aligned, and serious defects are detected. Save standard template files offline.

During online detection, the shape matching template is read, and the shape-based template matching algorithm is used to obtain the position information and angle information of the mobile phone screen in the image to be detected.

Use the threshold segmentation algorithm to segment the window area of the mobile phone screen in the image to be detected, and create an affine transformation matrix by obtaining the obtained position and angle information; read the standard area template of the window area, and use the area affine transformation to convert the read standard area The template is aligned with the window area to be detected; .

The aligned standard area template and the window area to be detected are checked for serious defects through Boolean operations.

Step 3: Extract the defects of the mobile phone screen image through a threshold segmentation algorithm, which specifically includes the following steps:

1), use median filter to preprocess the image to reduce noise.

As shown in FIG. 2 , the detection area of the picture, that is, the screen area, is divided into m×n areas (8×6 total 48 areas); due to the large size of the picture, the screen area is detected in blocks. On the one hand, block detection can avoid the problem of missed detection of defects caused by the global parameters of the algorithm. On the other hand, enabling multi-threading and parallel detection of each small block can effectively improve the detection efficiency.

Since the image size is very large, if some global parameters, such as global threshold, are directly used in the defect extraction process, it is easy to be affected by the problem of uneven illumination, which brings about the problem of missing defect detection. The imaging contrast of different defects is different, and some defects, such as shallow scratches, have very low imaging contrast and cannot be obtained directly through the threshold value, and need to be enhanced by algorithms such as image enhancement. In this embodiment, the Gaussian filter enhancement algorithm is used to perform Gaussian filtering on the image, and the filtered image is differentiated from the original image to achieve the effect of enhancing the defects to be detected. The enhancement effect can be controlled by the size of the filter template and the enhancement factor. It can be seen that Gaussian enhancement can perform local enhancement. Compared with Fourier transform and wavelet transform, the algorithm consumes less time and has better enhancement effect.

The image enhancement is performed on each area by using the Gaussian filter enhancement algorithm. The Gaussian filter enhancement algorithm is as follows:

img_enhanced=round((orig-mean_gauss)*factor)+orig(1)

In formula (1), mean_gauss represents the Gaussian filtering of the image, orig represents the corresponding gray value of each image, factor is the enhancement factor, and img_enhanced is the gray value of the enhanced output image; the round() function is a rounding function.

The enhanced image, the defect contrast is greatly improved.

4) Defects are extracted by threshold segmentation algorithm.

Step 4: As shown in Figure 3 and Figure 4, the clustering algorithm is used to connect the dense point clusters. Specifically, the cluster radius is R pixels, the area is greater than S pixels, and the number of small defect areas is greater than N All regions of each are clustered and connected, in this embodiment, R=50, S=15, and N=50. That is, in a circular area with a radius of 50 pixels, if the number of areas with an area larger than the set minimum area of 15 pixels is greater than 50, it is considered that this part of the area is a dense point cluster area. Use the morphological closing operation to connect the scattered points of this area into a large area.

Step 5: Extract the connected domain features of the defects, input the connected domain features into the neural network classifier, and use the neural network classifier to classify point-shaped, linear and planar defects, as shown in Figure 5.

The connected domain features are divided into basic features, shape features and geometric moment features; details are shown in Table 1.

Table 1 Connected Domain Features

Defect classification by the neural network classifier is mainly divided into two parts: offline training and online detection.

Before classifying the defects, the connected domain features of the defects extracted in step 5 are used as the defect data set; the connected domain features saved in the data set are divided into three categories, the first category is the basic features of area, center, width and height; The second category is the circumscribed circle radius, roundness, and compact shape features; the third category is the geometric moment features of the second and third order moments. In this embodiment, a total of 22-dimensional features are selected.

The neural network classifier includes an input layer, an intermediate layer and a hidden layer; the number of neurons in the input layer is the same as the dimension of the input feature vector, and the number of neurons in the output layer is the same as the classification output category. The hidden layer abstracts the features of the input data into another dimension space to show its more abstract features, which can be better linearly divided. Hidden layers can improve the accuracy and expressiveness of the network. As shown in Figure 6, it is a schematic diagram of the structure of the neural network classifier, which consists of 22 input layers, 10 hidden layers, and 2 output layers, and the activation function uses the ReLU activation function;

Offline training: set training parameters, the training parameters include optimization algorithm, learning rate, learning strategy, and initialization strategy. The optimization algorithm selects the momentum gradient descent method, the learning rate is 0.005 and the momentum coefficient is 0.9 according to experience. Weights and biases are initialized to small random numbers, such as a Gaussian distribution with mean 0 and variance 0.01. The number of iterations is 2000, the learning rate is changed every 500 times, and the iteration target is 0.1. Offline training is performed on the defect dataset, forward propagation calculates the loss function, and back propagation updates the weights. Until the loss function reaches the convergence state; save the trained neural network classifier offline;

Online detection: Extract the connected domain features of defects, read the trained neural network classifier, and input the extracted defect connected domain features into the classifier to obtain the defect classification results.

Step 6: Extract the area, length and radius of the defect area from the point-like, line-like and planar defect areas, and compare them according to the detection standard;

The area of the defect area is obtained by counting the number of pixels in the connected domain, which is used for the standard comparison of defects in the plane type; the length of the defect area is obtained by skeletonizing the defect area, which is used for the standard comparison of the length of the defect in the linear type; The radius of the defect area is the minimum circumcircle radius of the defect area, which is used for the comparison of defect radius standards.

According to the test standards, the test standards are as follows:

1) Point defects

D<0.1mm, ignore;

0.1mm≤D≤0.2mm, d≥10mm, if the number of defects within the radius of d is less than 2, it means that the mobile phone screen is a good product, and if the number of defects is greater than 2, it means that the mobile phone screen is a defective product;

D>0.2mm, NG means that the mobile phone screen is a defective product;

Among them, D is the diameter of the minimum circumscribed circle of point defects, and d is the distance between points;

2) Defective linearity

W<0.02mm, ignored, not a defect;

0.02mm≤W≤0.05mm, d≥5mm, L≤3, within 2 can be ignored, it is not a defect; within 3, NG indicates that the mobile phone screen is a defective product;

W>0.05mm or L>3mm, NG means the mobile phone screen is defective;

Among them, W is the width of the linear defect, L is the length of the linear defect, and d is the distance between different linear defects;

3) Poor face shape

D>0.5mm or S>0.2mm ²

Among them, D is the diameter of the largest inscribed circle of the planar defect, and S is the area of the planar defect.

Step 7: Use the deep learning classifier to reclassify the classified defects. Four types of defects were obtained by classification: scratches in Fig. 7a, floating filaments in Fig. 7b, light stains in Fig. 7c, and dark stains in Fig. 7d;

The deep learning classifier is a convolutional neural network model. As shown in Figure 8, the convolutional neural network model mainly includes an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer, wherein the input layer, Convolutional layer, pooling layer, convolutional layer, pooling layer, fully connected layer and output layer are connected in turn; The input, fully connected layer to the output layer completes the task of classification.

Defect classification by deep learning classifier is mainly divided into two parts: offline training and online detection.

Create a dataset of four types of defects: scratches, floating hairs, light stains, and dark stains. Enhance the dataset by mirroring, rotating, etc. In this embodiment, each type of original material is 500 pieces, and each type of defective material is 5000 pieces after being enhanced by 10 times.

A deep learning classifier based on convolutional neural network is used to set the model structure parameters and training parameters. When training offline, use the pre-trained model to perform transfer training on the dataset. After training, save the deep learning classifier offline. In this embodiment, the VGG16 convolutional neural network model of the Visual Geometry Group of Oxford University is used, which is composed of 13 convolutional layers, 5 pooling layers, and 3 fully connected layers. The GPU is used for migration training and the convolutional neural network pre-training model is used for migration training. The number of training iterations is 20,000, and every 32 pictures is a batch. The initial learning rate of the picture is 0.005, and the learning rate is changed every 5,000 times.

During online detection, the picture corresponding to the defect area to be classified is intercepted, and the picture is used as the input of the deep learning classifier to obtain the classification information of the defect. The CPU is used for testing, and the test image is changed to a single image for forward propagation calculation.

Step 8: Statistics of various types of defect information.

Through the rough inspection process, the present invention can firstly detect the serious defects in the window area, and through the fine inspection process, the classification of point-like, line-like and plane-like defects can be realized. Indistinguishable scratches, floating filaments, light stains and dark stains are also accurately identified. Figures 9 and 10 are schematic diagrams of the detection results.

The embodiments of the present invention are not limited by the above-mentioned examples, and any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principle of the present invention should be equivalent substitutions, and are included in this within the scope of protection of the invention.

Claims (10)

1. a mobile phone glass cover plate viewing window area defect detection method based on machine vision, is characterized in that, comprises the steps: Step 1. Collect mobile phone screen images; Step 2. Perform a rough inspection on the screen area of the mobile phone; Step 3: Extract the defects of the mobile phone screen image through a threshold segmentation algorithm; Step 4, using the clustering algorithm to connect the scattered points of the dense point cluster area on the mobile phone screen window area; Step 5: Extract the connected domain features of the defects, input the connected domain features into the neural network classifier, and use the neural network classifier to classify point-like, linear and planar defects; Step 6: Extract the area, length and radius of the defect area from the point-like, line-like and planar defect areas, and compare them according to the detection standard; Step 7. Use the deep learning classifier to reclassify the classified defects, and classify four types of defects: scratches, floating hairs, light-colored dirt, and dark-colored dirt; Step 8: Statistics of various types of defect information and quantity. 2. The method for detecting defects in the window area of a mobile phone glass cover plate based on machine vision according to claim 1, wherein the first step is to use a 16K line scan camera to perform image acquisition on the mobile phone screen, and the image resolution is 40000×16384 , the detection accuracy is 0.005mm. 3. The method for detecting defects in the window area of a mobile phone glass cover plate based on machine vision according to claim 1, is characterized in that, before carrying out the rough inspection, a picture of a defect-free good-quality mobile phone screen is selected, and the mobile phone screen area is framed. As the region of interest (ROI) of the standard template, the standard template file is saved offline; the standard template includes a shape matching template and a standard region template, wherein the shape matching template is mainly used for matching with the mobile phone screen in the image to be detected, and obtaining Obtain the position information of the mobile phone screen in the image to be inspected; the standard area template is mainly used to align with the area of the mobile phone screen extracted from the image to be inspected to detect defects; The rough inspection of the mobile phone screen area in step 2 includes the following steps: (1) During online detection, the shape matching template is read, and the shape-based template matching algorithm is used to match and obtain the position information and angle information of the mobile phone screen in the image to be detected; (2) Use the threshold segmentation algorithm to segment the window area of the mobile phone screen in the image to be detected, and create an affine transformation matrix by obtaining the obtained position and angle information; read the standard area template of the window area, and use the area affine transformation to read The standard area template is aligned with the window area to be detected; The aligned standard area template and the window area to be inspected are subjected to Boolean operation to detect whether there is a defect. 4. The method for detecting defects in the window area of a mobile phone glass cover plate based on machine vision according to claim 1, wherein the step 3 extracting the defect of the mobile phone screen image comprises the following steps: 1), use median filtering for image preprocessing to reduce noise; 2), dividing the window area to be detected, that is, the screen area, into m×n areas; 3), use the Gaussian filter enhancement algorithm to enhance the image of each segmented area. The Gaussian filter enhancement algorithm is: img_enhanced=round((orig-mean_gauss)*factor)+orig(1) In formula (1), mean_gauss represents the Gaussian filtering of the image, orig represents the corresponding gray value of each image, factor is the enhancement factor, and img_enhanced is the gray value of the enhanced output image; the round() function is a rounding function; 4), extract the defects in the window area of the mobile phone screen from the enhanced picture through the threshold segmentation algorithm. 5. The method for detecting defects in the window area of a mobile phone glass cover plate based on machine vision according to claim 1, wherein the clustering algorithm described in step 4 is to only connect the dense point regions that meet the clustering conditions, and not to The regions that meet the clustering conditions are not clustered and connected; The clustering conditions are: a. A circular dense point area with a clustering radius of R; b. In the circular area, the total number of areas where the area of the dense point area is greater than the area S of the minimum area of dense points is greater than the number of clusters N; In a circular area with a radius of R, if the number of areas that satisfy the requirement that the area of the dense point area is greater than the set minimum area S of the dense point area is greater than N, the dense point area is considered to be a dense point cluster area; using morphological expansion and The erosion operation connects the scatter points of this region into a region. 6. The method for detecting defects in the window area of mobile phone glass cover plates based on machine vision according to claim 1, wherein the neural network classifier described in step 5 is used for detecting point defects, linear defects and planar defects. Classification; The neural network classifier includes an input layer, an intermediate layer and a hidden layer; wherein, the number of neurons in the input layer is the same as the dimension of the input feature vector, and the number of neurons in the output layer is the same as the classification output category; The neural network classifier for defect classification is mainly divided into two parts: offline training and online detection; Before classifying the defects, the connected domain features of the defects extracted in step 5 are used as the defect data set; the connected domain features saved in the data set are divided into three categories, the first category is the basic features of area, center, width and height; The second category is the circumscribed circle radius, roundness and compact shape features; the third category is the geometric moment features of the second and third order moments; Offline training: Set training parameters, including optimization algorithm, learning rate, learning strategy, and initialization strategy, perform offline training on the defect data set, calculate the loss function through forward propagation, and update the weights through back propagation, until the loss function converges state; save the trained neural network classifier offline; Online detection: Extract the connected domain features of defects, read the trained neural network classifier, and input the extracted defect connected domain features into the classifier to obtain the defect classification results. 7. The method for detecting defects in the window area of a mobile phone glass cover plate based on machine vision according to claim 1, wherein the area of the defect area described in the step 6 is obtained by counting the number of pixels in the connected domain, using It is used for the standard comparison of defects in the plane type; the length of the defect area is obtained through the skeletonization of the defect area, which is used for the standard comparison of the defect length of the linear type; the radius of the defect area is the minimum circumcircle radius of the defect area, which is used for the defect radius Standard comparison. 8. The method for detecting defects in the window area of a mobile phone glass cover plate based on machine vision according to claim 1, wherein the detection standard of step 6 is: 1) Point defects D<0.1mm, ignore; 0.1mm≤D≤0.2mm, d≥10mm, if the number of defects within the radius of d is less than 2, it means that the mobile phone screen is a good product, and if the number of defects is greater than 2, it means that the mobile phone screen is a defective product; D>0.2mm, NG, indicating that the mobile phone screen is a defective product; Among them, D is the diameter of the minimum circumscribed circle of point defects, and d is the distance between points; 2) Defective linearity W<0.02mm, the defect is ignored; 0.02mm≤W≤0.05mm, d≥5mm, L≤3, if the defect is ignored within 2, it means that the mobile phone screen is a good product; if it is less than 3, it means that the mobile phone screen is a defective product; W>0.05mm or L>3mm, NG, indicating that the mobile phone screen is a defective product; Among them, W is the width of the linear defect, L is the length of the linear defect, and d is the distance between different linear defects; 3) Poor face shape D>0.5mm or S>0.2mm ² Among them, D is the diameter of the largest inscribed circle of the planar defect, and S is the area of the planar defect. 9. The method for detecting defects in the window area of a mobile phone glass cover plate based on machine vision according to claim 1, wherein the deep learning classifier is a convolutional neural network model, and the convolutional neural network model mainly comprises an input layer, convolutional layer, pooling layer, fully connected layer and output layer, among which, the input layer, convolutional layer, pooling layer, fully connected layer and output layer are connected in turn; the convolutional layer completes the feature extraction operation, and the last The defect features of the layer pooling layer are converted into a set of vectors as the input of the fully connected layer, and the defect features of the fully connected layer are input to the output layer to complete the classification task; Defect classification by deep learning classifier is mainly divided into two parts: offline training and online detection; Before classifying defects, create data sets of different types of defects, and use mirroring and rotation to enhance defect pictures in the data set; During offline training, use the pre-training model to perform migration training on the dataset; after training, save the deep learning classifier offline; During online detection, the image corresponding to the defect area to be classified is intercepted, and the image is used as the input of the deep learning classifier to classify the defects. 10. The method for detecting defects in the window area of a mobile phone glass cover plate based on machine vision according to claim 1, wherein a parameter configuration file is made to algorithm parameters and detection standards; the algorithm parameters include thresholds in the threshold segmentation algorithm, The number of blocks and the clustering radius in the clustering algorithm; the detection standard parameters are mainly the radius, length, width and number of defects involved in the point, line and plane detection standards.