CN114170140A

CN114170140A - A Yolov4-based Diaphragm Defect Recognition Method

Info

Publication number: CN114170140A
Application number: CN202111325709.5A
Authority: CN
Inventors: 甄志明; 卢清华; 陈勇
Original assignee: Foshan University
Current assignee: Foshan University
Priority date: 2021-11-10
Filing date: 2021-11-10
Publication date: 2022-03-11

Abstract

The invention provides a method for identifying diaphragm defects based on Yolov4, which comprises the following steps: firstly, acquiring diaphragm image data; secondly, marking the defect type by using a marking tool to generate a marking file as a data set; thirdly, clustering a prior frame of the data set by using a clustering algorithm k-means; fourthly, establishing an improved Yolov4 network model, and training the improved Yolov4 network model through a training set; the improved Yolov4 network model adopts a structure of combining low-level characteristic information and high-level characteristic information; CSP structures are added into SPP and PANet of the improved Yolov4 network model, and an attention mechanism is combined to improve the defect detection precision; fifthly, testing the detection performance of the trained improved Yolov4 network model by using a test set; and sixthly, deploying the training model of the optimal improved Yolov4 network model to a diaphragm detection site for diaphragm defect detection. The method has strong robustness, can reduce the omission factor and the false detection rate, and can improve the detection quality of the diaphragm.

Description

Membrane defect identification method based on Yolov4

Technical Field

The invention relates to the technical field of diaphragm detection, in particular to a method for identifying diaphragm defects based on Yolov 4.

Background

The lithium battery is visible everywhere in daily life, such as a mobile phone, a battery car, a new energy automobile and the like which are commonly used. The demand is increasing, and lithium cell industry is also continuous to put into production, but some problems are also emerging gradually, for example, a certain brand cell-phone charges and explodes, or the battery has the problem such as electric leakage. The safety and durability of the battery need to be further enhanced. The diaphragm is an important component of the lithium battery, isolates the positive electrode and the negative electrode of the battery and provides a flow passage for the battery.

However, in an automatic production line, the membrane is inevitably subjected to some collisions, folding and friction, and the surface of the membrane may have creases, scratches and pinholes, or factors such as missing spraying or less spraying in the spraying process of the membrane. As a raw material of the battery, the quality of the separator also affects the quality and safety of the battery, and therefore, it is necessary to check the production quality of the separator.

At the present stage, the manual diaphragm detection mode is gradually eliminated, and the detection efficiency cannot meet the production requirements of modern industry. With the continuous progress of cameras and image algorithms, visual inspection is increasingly applied to the industry. Since the diaphragm is produced at first and is subjected to rolling and splitting procedures, the diaphragm is conveyed and photographed on a conveyor belt for detection, and some manufacturers choose to use a line-scan camera to acquire an image of the diaphragm and use traditional image algorithms such as morphological processing, image segmentation and the like to detect the quality of the diaphragm. However, in an automatic production mode, the conventional image processing algorithm has a considerable effect on detection of defect features, but the defects on the diaphragm have different forms, and gray values of some defects are similar to the background of a normal diaphragm, so that the conventional image processing algorithm has a high probability of missing detection and false detection in this respect, and meanwhile, the scheme design of the conventional image processing algorithm is also largely related to the technical level of technicians, and the production environment also affects the detection effect.

Disclosure of Invention

The invention aims to overcome the defects and shortcomings in the prior art, and provides a method for identifying diaphragm defects based on Yolov4, which has strong robustness, can reduce the omission factor and the false detection factor, and can solve the problem that the traditional algorithm cannot detect the fine defects of the diaphragm, thereby achieving better positioning and classifying effects on the diaphragm defects and further improving the detection quality of the diaphragm.

In order to achieve the purpose, the invention is realized by the following technical scheme: a method for identifying diaphragm defects based on Yolov4 is characterized in that: the method comprises the following steps:

firstly, acquiring diaphragm image data, and screening defective pictures from the image data;

secondly, marking the defect type by using a marking tool to generate a marking file as a data set;

thirdly, clustering a prior frame of the data set by using a clustering algorithm k-means;

fourthly, dividing the data set into a training set, a verification set and a test set; establishing an improved Yolov4 network model, and training the improved Yolov4 network model through a training set; the improved Yolov4 network model adopts a structure of combining low-level characteristic information and high-level characteristic information; CSP structures are added into SPP and PANet of the improved Yolov4 network model, and an attention mechanism is combined to improve the defect detection precision;

fifthly, testing the detection performance of the trained improved Yolov4 network model by using the test set, and using the optimal training model of the improved Yolov4 network model for detection;

and sixthly, deploying the training model of the optimal improved Yolov4 network model to a diaphragm detection site for diaphragm defect detection, decoding the prediction result, performing score sorting and non-maximum inhibition screening on the prediction result, and drawing the processed result on the diaphragm original image.

In the first step, after defective pictures are screened from image data, the defective pictures are cut randomly; and rotating and folding the cut image to increase sample data.

Clockwise 90 DEG, 180 DEG and 270 DEG rotation is carried out on the cut image; and horizontally folding and vertically folding the cut image. The method can enhance the sample data to expand the data set, avoid overfitting, simultaneously improve the generalization ability of the subsequent model, and enhance the diversity of the sample by using the data.

In the second step, the defect type is marked by using a marking tool, and a marking file is generated as a data set, wherein the marking file is as follows: and manually labeling the image with the diaphragm defects by using a labelImg picture labeling tool, labeling the types of the defects and the minimum external rectangular frame of the defects, and generating an xml file format for storing the sizes and the types of the pictures of the defects and the coordinate information of the labeled rectangular frame as a data set.

The improved Yolov4 network model comprises a module 1, a module 2, a module 3 and a Head network Yolo Head; the module 1 is a trunk feature extraction network COSA-2x2 x; the module 2 adds a CSP structure on the basis of SPP; the module 3 adds a CSP structure to the up-sampling and down-sampling of the PANET structure and adds an attention mechanism.

And adding a feature fusion layer on the top feature layer of the module 3, wherein the feature fusion layer comprises 3 convolution layers, and the combination of low-layer feature information and high-layer feature information is realized.

The CSP structure firstly divides an input part into two parts, then one part passes through a Conv block part of a main body, and the other part is directly stacked with an output after the Conv block; wherein, dividing the input part into two parts means that the input part is divided into two parts on average, and the number of channels is half of the original input; or dividing the input part into two means performing convolution by 1 × 1, and halving the number of channels.

The fourth step includes the steps of:

dividing a data set into a training set, a verification set and a test set according to the proportion of 0.7:0.15: 0.15;

step two, improving a Yolov4 network model for initialization;

thirdly, obtaining an output value of the input training set data through a backbone feature extraction network COSA-2x2x, a module 2, a module 3 and a Head network Yolo Head;

step four, solving the error between the output value of the improved Yolov4 network model and the target value, namely a loss function;

and step five, updating the weight, and finishing the training when the improved Yolov4 network model converges to a certain degree and does not drop any more.

In the third step, the training set data extracts the defect characteristic information of the diaphragm through a trunk characteristic extraction network COSA-2x2x, the receptive field is increased through a module 2, the context information is separated, the characteristics are repeatedly extracted through a module 3, and finally the obtained characteristics are predicted through a Head network Yolo Head.

The loss function in step four is

loss＝loss_{Boundary frame}+loss_{Confidence level}+loss_{Classification}；

loss_{Boundary frame}＝loss_CDIoU＝1-IoU+ρ²(box_dt,box_gt)/c²+αν；

The CSPdacrnet 53 has strong feature extraction capability without using the original trunk network CSPdacrnet 53 and CSPdacrnet 53 of yolov4, but a large amount of display card memories occupied by calculation and network training are correspondingly brought, and the detection speed is slower than that of COSA-2x2 x. The COSA-2x2x with the PCB technology can make the model more flexible, and the speed is greatly improved. In addition, the CSP structure added on the basis of the SPP can effectively reduce the calculation amount of the model under the condition of little precision loss.

CSP structures are added to the up-sampling and down-sampling of the PANet structure, and an attention mechanism is integrated into the PANet structure, so that an improved Yolov4 network model can pay more attention to an ROI (region of interest) in a defect, the key information of the defect of the diaphragm is extracted, most irrelevant background information of the diaphragm of the lithium battery is ignored, the number of the parameters can be reduced, and the detection precision is improved. Because the diaphragm has some tiny defects, the invention adds a feature fusion layer on the top feature layer of the module 3 to realize the combination of low-layer feature information and high-layer feature information and carry out subsequent defect detection and positioning, so that the output feature layers are (104, 104 and 128), the micro-miniature defect detection device is more suitable for detecting micro-miniature targets, and has higher positioning precision and reduced false detection rate of tiny defects.

Compared with the prior art, the invention has the following advantages and beneficial effects: the method for identifying the diaphragm defects based on the Yolov4 has strong robustness, can reduce the omission factor and the false detection rate, and can solve the problem that the traditional algorithm cannot detect the fine defects of the diaphragm, thereby achieving better positioning and classifying effects on the diaphragm defects and further improving the detection quality of the diaphragm.

Drawings

FIG. 1 is a flow chart of a Yolov 4-based method for identifying diaphragm defects in accordance with the present invention;

FIG. 2 is a schematic diagram of the improved Yolov4 network model of the present invention;

FIG. 3 is a CSP structure diagram in the improved Yolov4 network model according to the present invention;

FIG. 4 is a diagram of CSP structure and 5 convolutions in the improved Yolov4 network model according to the present invention;

5(a) -5(e) are graphs showing the detection effect of the method for identifying the defect of the diaphragm based on Yolov 4;

Detailed Description

The present invention will be described in further detail with reference to the accompanying drawings and specific embodiments.

Examples

As shown in fig. 1 to 5(e), the method for identifying a defect of a diaphragm based on Yolov4 of the present invention includes the following steps:

firstly, acquiring diaphragm image data, and screening defective pictures from the image data; after screening defective pictures from the image data, randomly cutting the defective pictures; and rotating the cut image by 90 degrees, 180 degrees and 270 degrees clockwise, and performing horizontal folding and vertical folding to increase sample data.

And secondly, manually labeling the image with the diaphragm defects by using a labelImg picture labeling tool, labeling the types of the defects and the minimum external rectangular frame of the defects, and generating an xml file format for storing the sizes and the types of the pictures of the defects and the coordinate information of the labeled rectangular frame as a data set.

And thirdly, clustering 9 groups of prior frames of the data set by using a clustering algorithm k-means, and taking the 9 groups of prior frames as an anchor box of the improved Yolov4 network model in the fourth step, so that the prior frames are more in line with the size and distribution condition of the rectangular frame of the diaphragm defect.

sixthly, deploying a training model of an optimal improved Yolov4 network model to a diaphragm detection site to perform diaphragm defect detection, decoding a prediction result, performing score sorting and non-maximum inhibition screening on the prediction result, and drawing a processing result on a diaphragm original image, wherein fig. 5(a) -5(e) are effect images for respectively detecting creases, black spots, missed spraying, scratches and pinholes by adopting the Yolov 4-based diaphragm defect identification method.

Specifically, the improved Yolov4 network model comprises a module 1, a module 2, a module 3 and a Head network Yolo Head, wherein the module 1 is a backbone feature extraction network COSA-2x2x, the module 2 is a CSP structure added on the basis of SPP, and the module 3 is a CSP structure added on the upsampling and downsampling of a PANet structure and an attention mechanism is added. According to the invention, a feature fusion layer is added on the top feature layer of the module 3, and the feature fusion layer comprises 3 convolution layers, so that the combination of low-layer feature information and high-layer feature information is realized.

As shown in fig. 3, the CSP structure first divides the input part into two parts, then one part passes through the trunk Conv block part, and the other part is directly stacked with the output after passing through the Conv block; wherein, dividing the input part into two parts means that the input part is divided into two parts on average, and the number of channels is half of the original input; or dividing the input part into two means performing convolution by 1 × 1, and halving the number of channels.

The CSP structure is added to the up-sampling and the down-sampling of the PANet structure, the attention mechanism is integrated into the PANet structure, an improved Yolov4 network model is enabled to pay more attention to an ROI (region of interest) in a defect, key information of the defect of the diaphragm is extracted, most irrelevant background information of the diaphragm of the lithium battery is ignored, the number of the parameters can be reduced, and the detection precision can be improved, wherein the CSP structure and a 5-time convolution schematic diagram are shown in figure 4. Because the diaphragm has some tiny defects, the invention adds a feature fusion layer on the top feature layer of the module 3 to realize the combination of low-layer feature information and high-layer feature information and carry out subsequent defect detection and positioning, so that the output feature layers are (104, 104 and 128), the micro-miniature defect detection device is more suitable for detecting micro-miniature targets, and has higher positioning precision and reduced false detection rate of tiny defects.

The Yolo Head of the present invention uses the extracted features to perform prediction, and predicts 3 feature scales including width, height and channel number, which are (104, 104, 128), (26, 26, 512), (13, 13, 1024), respectively. All convolutions with a convolution kernel of 1x1 of the present invention use the Mish activation function. Mish activation function has better smoothness than Leaky-ReLU, and can allow better information to be transmitted into a neural network, so that better accuracy and generalization capability are obtained. The expression of the Mish activation function is as follows:

Mish＝x×tanh(In(1+e^x))

the fourth step of the present invention comprises the steps of:

step one, dividing a data set into a training set, a verification set and a test set according to the proportion of 0.7:0.15: 0.15.

And step two, initializing the improved Yolov4 network model.

And step three, extracting defect characteristic information of the diaphragm by the training set data through a trunk characteristic extraction network COSA-2x2x, increasing the receptive field through a module 2, separating context information, repeatedly extracting characteristics through a module 3, and finally predicting the obtained characteristics through a Head network Yolo Head.

Step four, solving the error between the output value of the improved Yolov4 network model and the target value, namely a loss function; wherein the loss function is

loss＝loss_{Boundary frame}+loss_{Confidence level}+loss_{Classification}；

loss_{Boundary frame}＝loss_CDIoU＝1-IoU+ρ²(box_dt,box_gt)/c²+αν；

The detection performance of the Yolov 4-based diaphragm defect identification method adopted in the present embodiment is shown in table 1 below.

TABLE 1 test performance of the model

	Speed of detection (ms/piece)	Recall (%)	Model parameter quantity params
				Original model algorithm	54	88.13	64,040,001
The example method	28	91.80	15,178,913
				SSD	20	48.18	6,204,004
YOLOv4-tiny	9	75.35	5,961,014

Wherein, the recall ratio formula is as follows:

the True Positives (TP) indicates that it is predicted to be a positive sample, actually a positive sample. False Positive (FP) indicates a predicted positive sample, and actually a negative sample. The higher the recall, the higher the probability that the actual defect is predicted.

The above embodiments are preferred embodiments of the present invention, but the present invention is not limited to the above embodiments, and any other changes, modifications, substitutions, combinations, and simplifications which do not depart from the spirit and principle of the present invention should be construed as equivalents thereof, and all such changes, modifications, substitutions, combinations, and simplifications are intended to be included in the scope of the present invention.

Claims

1. a diaphragm defect identification method based on Yolov4, is characterized in that: comprise the following steps:

The first step is to obtain diaphragm image data, and screen defective pictures from the image data;

The second step is to use the labeling tool to label the defect type, and generate a labeling file as a data set;

The third step is to use the clustering algorithm k-means to cluster the a priori frame of the data set;

The fourth step is to divide the data set into training set, verification set and test set; establish an improved Yolov4 network model, and train the improved Yolov4 network model through the training set; among them, the improved Yolov4 network model adopts low-level feature information and high-level feature information. Combined structure; the CSP structure is added to the SPP and PANet of the improved Yolov4 network model, and combined with the attention mechanism to improve the detection accuracy of defects;

The fifth step is to use the test set to test the detection performance of the improved Yolov4 network model after training, and use the optimal training model of the improved Yolov4 network model for detection;

The sixth step is to deploy the training model of the optimal improved Yolov4 network model to the diaphragm inspection site for defect detection of the diaphragm, and decode the prediction results. Draw on the original image.

2. the diaphragm defect identification method based on Yolov4 according to claim 1, is characterized in that: in the first step, after screening the defective picture from the image data, the defective picture is randomly cropped; The image is rotated and folded to increase the sample data.

3. the diaphragm defect identification method based on Yolov4 according to claim 2, is characterized in that: 90 °, 180 ° and 270 ° of rotations clockwise are carried out to the cropped image; the cropped image is horizontally folded and vertically Fold over.

4. the diaphragm defect identification method based on Yolov4 according to claim 1, is characterized in that: in the second step, described use labeling tool to label defect type, generate labeling file, refer to as data set: use labelImg picture labeling tool Manually mark the image with diaphragm defects, mark the category of the defect and the minimum bounding rectangle of the defect, and generate the .xml file format to store the image size, category and coordinate information of the marked rectangle as data. set.

5. the diaphragm defect identification method based on Yolov4 according to claim 1, is characterized in that: described improved Yolov4 network model comprises module 1, module 2, module 3 and head network Yolo Head; Described module 1 is backbone feature Extract the network COSA-2x2x; the module 2 adds a CSP structure on the basis of SPP; the module 3 adds a CSP structure to the upsampling and downsampling of the PANet structure, and adds an attention mechanism.

6. the diaphragm defect identification method based on Yolov4 according to claim 5, is characterized in that: at the topmost feature layer of module 3, add feature fusion layer, this feature fusion layer is 3 convolution layers, realizes low-level feature information and. High-level feature information is combined.

7. the diaphragm defect identification method based on Yolov4 according to claim 5, is characterized in that: CSP structure first divides the input part into two parts, then part passes through the backbone Conv block part, and another part directly stacks with the output after passing through the Convblock ; Among them, dividing the input part into two parts means splitting the input part into two parts on average, and the number of channels is half of the original input; or dividing the input part into two parts means performing a 1×1 convolution to halve the number of channels.

8. the diaphragm defect identification method based on Yolov4 according to claim 6, is characterized in that: the 4th step may further comprise the steps:

Step 1, the data set is divided into training set, validation set and test set with a ratio of 0.7:0.15:0.15;

Step 2, improve the Yolov4 network model for initialization;

Step 3, the input training set data obtains the output value through the backbone feature extraction network COSA-2x2x, module 2, module 3 and the head network Yolo Head;

Step 4: Find the error between the output value of the improved Yolov4 network model and the target value, that is, the loss function;

In step 5, the weights are updated. When the improved Yolov4 network model converges to a certain extent and no longer decreases, the training ends.

9. the diaphragm defect identification method based on Yolov4 according to claim 8, is characterized in that: in step 3, training set data extracts the defect feature information of diaphragm by backbone feature extraction network COSA-2x2x, increases receptive field through module 2, The context information is separated, and the features are repeatedly extracted by module 3, and finally the obtained features are predicted by the head network Yolo Head.

10. the diaphragm defect identification method based on Yolov4 according to claim 8, is characterized in that: the loss function in the step 4 is

loss=loss _{bounding box} +loss _confidence +loss _{classification} ;

loss _{bounding box} =loss _CDIoU =1-IoU+ρ ² (box_dt,box_gt)/c ² +αν;