CN114663346A - Strip steel surface defect detection method based on improved YOLOv5 network - Google Patents

Abstract

The invention discloses a strip steel surface defect detection method based on an improved YOLOv5 network, which is based on a YOLOv5 network model and is added with a self-designed channel space attention module, thereby improving the detection precision and solving the problem of feature extraction in complex scenes and backgrounds. The detection method provided by the invention gives full play to the advantage of extracting the characteristics by the deep learning method, can learn simple shallow characteristics from a large amount of data set without depending on manual characteristic engineering, and then gradually learn more complex and abstract deep characteristics, and has the advantages of better performance, higher defect type identification precision, high defect accuracy and recall rate of the lithium battery and high identification speed.

Description

A strip surface defect detection method based on improved YOLOv5 network

technical field

The invention belongs to the technical field of industrial defect detection, and in particular relates to a strip steel surface defect detection method based on a YOLOv5 network of a spatial channel attention module.

Background technique

Strip steel is one of the important raw materials of steel, widely used in machinery manufacturing, aerospace and transportation, and plays an important role in various production and life. It may cause various defects such as surface oil spots, almond-shaped defects, white spots and scratches on the surface of the strip. These defects greatly affect the corrosion resistance and service life of the strip. The existing defect detection methods mainly use artificial naked-eye detection, which has low detection efficiency, high labor intensity and high production cost, and can no longer meet the needs of strip surface defect detection.

Deep learning automatically extracts and learns defect features through convolutional neural networks, without the need for artificial feature factor design. Therefore, deep neural networks have the characteristics of strong learning ability and high robustness, and have gradually become the mainstream of strip surface defect detection. method. Weng Yushang et al. (Weng Yushang, Xiao Jinqiu, Xia Yu. Improved Mask R-CNN Algorithm for Strip Steel Surface Defect Detection [J/OL]. Computer Engineering and Applications: 1-12 [2021-06-24].) Proposed an improved mask Code Region Convolutional Neural Network (Mask R-CNN) algorithm, using k-means II clustering algorithm to improve the region proposal network (RPN) anchor box generation method. Li Weigang et al. (Li Weigang, Ye Xin, Zhao Yuntao, Wang Wenbo. Strip surface defect detection based on improved YOLOv3 algorithm [J]. Journal of Electronics, 2020) proposed a YOLOV3 algorithm framework that combines shallow features and deep features, The improved YOLOv3 algorithm achieves an average accuracy of 80% on the strip steel dataset of Northeastern University. However, in the face of weak and tiny strip defects, due to the strong coupling between the background and the foreground and the small defect area, the deep learning model cannot extract features well, resulting in poor model detection effect.

SUMMARY OF THE INVENTION

In view of the deficiencies of the prior art, the technical problem to be solved by the present invention is to provide a strip steel surface defect detection method based on an improved YOLOv5 network, which can perform real-time detection and defect positioning for common different types of strip steel surface defects, and improve the The accuracy of identification of different types and similar structural defects can meet the real-time and accuracy requirements of actual strip steel industrial production.

The technical solution adopted by the present invention to solve the above-mentioned technical problems is to design a method for detecting surface defects of strip steel based on an improved YOLOv5 network, which is characterized in that the method comprises the following steps:

Step 1: Image dataset acquisition

1.1 Use an industrial camera to collect images of the surface of the strip steel, and screen out the pictures containing defects; when the defect types in the screened defect images cover the types of known surface defects of the strip steel, a defect picture set is formed;

1.2 Normalize the size of the defect picture set, and then use Labelimg software to manually mark the pictures in the defect picture set, so that each defect picture has the label of the defect type and the coordinates of the defect position;

1.3 Randomly divide no less than 60% of the marked defect image set into the training set, and the rest as the verification set;

Step 2: Construct an improved YOLOv5 network model

The improved YOLOv5 network model is based on the YOLOv5 network model, and a CSA module is connected in series between the three CSP23_ modules of the PAN and the three conv modules of the classification and positioning part of the network model;

The CSA module includes a channel attention module and a spatial attention module, the two modules are connected in series, and the output of the channel attention module is the input of the spatial attention module;

The feature map F1 output by the _{CSP23_} module is input into the CSA module, and firstly processed by the channel attention module. The channel attention module _first passes the input feature F1 through global max pooling and global average pooling based on depth and width, respectively. , two 1×1×C feature maps are obtained; then, the two obtained 1×1×C feature maps are processed by a fast one-dimensional convolution with a convolution kernel size of k, and then the two fast The results obtained by the one-dimensional convolution are added and then processed by the activation function sigmoid to obtain the channel attention; the channel attention is multiplied by the original feature F ₁ to re-weight the feature to obtain the weighted feature F ₂ ;

_The feature F2 output by the channel attention module is input to the spatial attention module, and the spatial attention module performs global maximum pooling and global average pooling on the feature F2 respectively, and obtains _two feature maps of H×W×1; then The two H×W×1 feature maps are spliced based on the channel, and then the result obtained by the splicing operation is subjected to a convolution operation with a convolution kernel of 7×7 size, and the dimension is reduced to a one-dimensional vector, that is, H×W×1, and then generate the spatial attention weight through the activation function sigmoid; finally, multiply the spatial attention weight with the input feature F ₂ to obtain the output feature F ₃ of the spatial attention module; feature F ₃ is the CSA module’s output feature F 3 . The output is also the input of the conv module of the classification and positioning part of the network model;

Step 3: Train and improve the YOLOv5 network model

3.1 Image dataset preprocessing

The training set is preprocessed by Mosaic data augmentation;

3.2 Parameter setting

Initialize all weight values, bias values, and batch normalized scale factor values, set the initial learning rate and batch_size of the network, and input the initialized parameter data into the network; dynamically adjust the learning rate and the number of iterations according to changes in training loss, In order to update the parameters of the entire network; the training is divided into two stages, the first stage is the first 100 cycles of training, and the initial learning rate is fixed at 0.001 to speed up the convergence; the second stage refers to the training cycle after 100 cycles, The initial learning rate is set to 0.0001;

3.3 Network model training

Input the preprocessed training set into the improved YOLOv5 network model after the initialization parameters are set in the second step for feature extraction, and use the K-means clustering method to automatically generate anchor frames for the training set images. The size is used as the prior box, and the bounding box is obtained by bounding box regression prediction; then the bounding box is classified by the logistic classifier to obtain the classification probability of the defect category corresponding to each bounding box; The classification probability of defect categories is sorted, the defect category corresponding to each bounding box is determined, and the predicted value is obtained; the predicted value includes defect category and defect location information, and the non-maximum suppression threshold is 0.5; then the predicted value and the true value are calculated by the loss function GIOU The loss value between them; backpropagation is performed according to the training loss value, and the parameters of the backbone network and the classification and regression network are updated, until the loss value conforms to the preset value, and the training of the network model parameters is completed;

3.4 Network Model Test

Input the validation set into the network model that has completed the parameter training in step 3.3, and obtain the tensor prediction value of the validation set; compare the tensor prediction value with the label information to test the reliability of the network model; Evaluation, when the AP is not less than 85%, the network model test is reliable;

Step 4: Strip surface defect detection

The surface image of the strip to be inspected is subjected to the same size normalization operation as in step 1.2 of the first step, and then input into the network model tested as reliable in the third step to obtain the defects of the surface image of the strip to be inspected Tensor information, including defect location, defect category, and confidence.

Compared with the prior art, the beneficial effects of the present invention are: the detection method of the present invention is based on the YOLOv5 network model, and a self-designed channel space attention module is added; Integrate them together and then perform the attention operation, so that since the deep layer contains more high-level semantic information and less background information, the target information after the fusion of shallow features and deep features is strengthened, and the network can be made when doing attention operations. More attention is paid to target defects and background information is suppressed, which can better guide multi-scale fusion, improve detection accuracy, and solve the problem of feature extraction in complex scenes and backgrounds. The detection method of the present invention fully utilizes the advantages of the deep learning method for extracting features, and can learn simple shallow features from a large number of data sets without relying on artificial feature engineering, and then gradually learn more complex and abstract deep features, with better performance. , the defect type identification accuracy is higher, and the lithium battery defect has a high precision rate, a high recall rate, and a fast identification speed.

Description of drawings

FIG. 1 is a schematic diagram of the structure and principle of a CSA module according to an embodiment of the detection method of the present invention.

2 is a schematic diagram of the structure and principle of an improved YOLOv5 network model according to an embodiment of the detection method of the present invention.

Detailed ways

The technical solutions of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative work fall within the scope of protection of the present application.

The present invention provides a kind of strip steel surface defect detection method (referred to as detection method, referring to Fig. 1-2) based on improved YOLOv5 network, this detection method comprises the following steps:

Step 1: Image dataset acquisition

1.1 Use an industrial camera to collect images of the surface of the strip steel, and screen out the pictures containing defects; when the defect types in the screened defect images cover the types of known surface defects of the strip steel, a defect picture set is formed;

1.2 Normalize the size of the defect picture set (scale to 608*608 pixels in this embodiment), and then use Labelimg software to manually mark the pictures in the defect picture set, so that each defect picture has defect types and defects. label for location coordinates;

1.3 Randomly divide no less than 60% of the marked defect image set into the training set and the rest as the validation set; this implementation is 4:1, that is, the training set is 80%, and the remaining 20% is the validation set.

Step 2: Construct an improved YOLOv5 network model

The improved YOLOv5 network model is based on the YOLOv5 network model, in the three CSP23_ (cross-stage local networks, respectively CSP23_5, CSP23_4, CSP23_3) modules of the PAN (pixel aggregation network) and three of the classification and positioning parts of the network model. A CSA (channel spatial attention) module is connected in series between the conv (convolution) modules.

The CSA module includes a channel attention module and a spatial attention module. The two modules are connected in series, and the output of the channel attention module is the input of the spatial attention module.

The feature map F ₁ (C×W×H) output by the CSP23_ module is input into the CSA module, and firstly processed by the channel attention module, which firstly passes the input feature F ₁ (C×W×H) through the Based on depth and width global max pooling (MaxPool) and global average pooling (AvgPool), two feature maps of C×1×1 are obtained. Next, the obtained two C×1×1 feature maps are processed by a fast one-dimensional convolution (Conv1d) with a convolution kernel size of k, and then the results obtained by the two fast one-dimensional convolutions are added. Then, through the activation function sigmoid processing, the channel attention is obtained; the channel attention is multiplied by the original feature F ₁ to re-weight the feature, and the weighted feature F ₂ is obtained.

The kernel size k of fast 1D convolution represents the coverage of local cross-channel interactions, that is, how many neighbors participate in the attention prediction of 1D channels. The coverage of the interaction (that is, the kernel size k) is proportional to the channel dimension, and the specific calculation formula is:

where C represents the number of feature channels, and β=2 and b=1 represent two hyperparameters.

_The feature F2 output by the channel attention module is input to the spatial attention module, and the spatial attention module performs global maximum pooling (Max Pool) and global average pooling (Mean Pool) on the feature F2 respectively, and obtains _two 1× The feature map of W×H; then the two 1×W×H feature maps are concatenated based on the channel (Concat) operation, and then the result obtained by the concatenation operation is passed through a convolution kernel with a size of 7×7. Operation, reduce the dimension to a one-dimensional vector, that is, 1×W×H, and then generate the spatial attention weight through the activation function sigmoid. Finally, the spatial attention weight is multiplied by the input feature F ₂ to obtain the output feature F ₃ of the spatial attention module. Feature F3 is the output _of the CSA module and the input of the conv module of the classification and localization part of the network model.

Step 3: Train and improve the YOLOv5 network model

3.1 Image dataset preprocessing

The training set is preprocessed by Mosaic data augmentation;

3.2 Parameter setting

Initialize all weight values, bias values, and batch normalized scale factor values, set the initial learning rate and batch_size of the network, and input the initialized parameter data into the network; dynamically adjust the learning rate and the number of iterations according to changes in training loss, In order to update the parameters of the entire network; the training is divided into two stages, the first stage is the first 100 cycles of training, and the initial learning rate is fixed at 0.001 to speed up the convergence; the second stage refers to the training cycle after 100 cycles, The initial learning rate is set to 0.0001.

3.3 Network model training

Input the preprocessed training set into the improved YOLOv5 network model after the initialization parameters are set in the second step for feature extraction, and use the K-means clustering method to automatically generate anchor frames for the training set images. The size (the anchor box size will be proportionally scaled according to the image zoom size) is used as the a priori box, and the bounding box is obtained through the bounding box regression prediction; then the logistic classifier is used to classify the bounding box to obtain the defect category classification probability corresponding to each bounding box; Then, the non-maximum value suppression method (NMS) is used to sort the defect category classification probabilities of all bounding boxes, determine the defect category corresponding to each bounding box, and obtain the predicted value; the predicted value includes the defect category and defect location information, and the non-maximum The suppression threshold is 0.5; then the loss value between the predicted value and the real value is calculated by the loss function GIOU; back-propagation is performed according to the training loss value, and the parameters of the backbone network and the classification and regression network are updated until the loss value conforms to the preset, the network model The training of parameters is completed;

3.4 Network Model Test

Input the validation set into the network model that has completed the parameter training in step 3.3, and obtain the tensor prediction value of the validation set; compare the tensor prediction value with the label information to test the reliability of the network model; Evaluation, when the AP is not less than 85%, the network model test is reliable.

Step 4: Strip surface defect detection

The surface image of the strip to be inspected is subjected to the same size normalization operation as in step 1.2 of the first step, and then input into the network model tested as reliable in the third step to obtain the defects of the surface image of the strip to be inspected Tensor information, including defect location, defect category, and confidence (maximum probability of defect category classification).

The YOLOv5 model uses the Mosaic data enhancement method for the input image to enrich the image content and improve the detection effect of light smoke or small smoke area.

Mosaic data augmentation has the following characteristics:

First, take a batch of images from the training sample set (batch refers to Batch, batch refers to Batch size, which is a hyperparameter of the model, in this example, the batch is equal to 32) images; secondly, 4 images are randomly selected from this batch of images, Randomly operate the four images through color gamut change, reduction, inversion and/or cropping, and perform at least one operation on each image; then arrange the four images in four directions: upper left, lower left, upper right, and lower right After stitching, a new image is obtained. The size of the new image is the same as the size of the original image without operation, which is 608×608×3; repeat the above operation, and set the number of cycles equal to the Batch size.

The Focus module connects three groups of CBL (Convolution-Batch Normalization-Batch Normalization-Activation Function Leaky Relu, CBL) convolution structure and a structure composed of cross-stage local network CSP1_X modules, and then connects a pooling network SPP module to form features Extract the network; by using the Focus module to slice the image, after passing through the Focus module, it is sent to the structure composed of the CBL convolution structure and the CSP1_X module, and then the low-level features and high-level features are fused through the SPP module.

The Focus module has the following features:

First mark the original image of size 608×608×3 with four numbers 1, 2, 3, and 4, and then combine the pixels of the same number into 4 parts of size 304×304×3, and then combine these 4 Each part is spliced into a 304×304×12 feature map in the depth direction according to the number size, and then connected to a CBL convolution structure.

The features of the CBL convolution structure included in the Focus module are: the number of convolution kernels in the convolution layer (convolution, conv) is 64, the size is 3 × 3, and the stride is 1.

The CSP1_X module has the following characteristics:

X represents the number of residual structures, and other structures are the same except for the number of residual structures.

Taking the CSP1_3 module as an example, the CSP1_3 module first performs the CBL convolution operation on the input feature map; The new feature map obtained after direct convolution is concatenated in the depth direction; finally, it is input into the next module through batch normalization, activation function Leakyrelu and a layer of CBL convolution structure.

The convolution layer conv of the CBL convolution structure that directly convolves the input feature map in the CSP1_X module is the same as the convolution layer conv located in the last CBL convolution structure in the CSP1_X module. The relevant dimensions of the convolution layer conv are: convolution The kernel size is 1×1 and the stride is 1.

Feature reprocessing network design: adopt the structure of feature pyramid FPN and pixel aggregation network PAN, where the FPN structure is composed of two sets of CSP2_3 modules, CBL convolution structure and up-sampling up sample structure in series; PAN includes two CBL convolution structures , downsample the data.

Concat the output of each upsampling structure in the FPN and the output feature maps of the CSP1_9-1 and CSP1_9-2 modules in the feature extraction network for depth-wise tensor splicing concat, and concat the output of each CBL convolution structure in the FPN Perform depthwise tensor splicing concat with the feature map of the corresponding size of the CBL convolution structure in the PAN. After each time the PAN structure passes through the CBL convolution structure, a CSP2_3 module and an SPP module are added respectively. The first CBL convolution structure of the PAN structure A CSP2_3 module and an SPP module were also added before; neither the CSP2_3 nor the SPP module changed the feature map size. The output of the CSP2_3-3 module in Neck was 76×76, and then connected to the CBL convolution structure to change the image size to 38×38 , the output of the CSP2_3-4 module is connected to the input of the CBL convolutional structure, turning the image into 19×19.

FPN includes two CBL convolution structures, but the CBLs here are convolutions with a stride of 1, which has no effect on the size of the feature map. It is the size of the feature map changed by the up-sampling up sample in the FPN. Learning multi-scale targets information.

The features of the cross-stage local network CSP2_3 module are:

The structure of this module is the same as the structure after deleting the add fusion process in each residual structure of the CSP1_3 module. The CSP2_3 module consists of an initial CBL convolutional structure and a final CBL convolutional structure, multiple repeating units, a convolutional layer of size 1×1 connected to a 3×3 convolutional layer by batch normalization and activation function Leaky relu. The 3×3 convolutional layer is then connected to the batch normalization and activation function Leaky relu to form a repeating unit. The number of repeating units is three. The input of the first repeating unit is connected to the output of the initial CBL convolution structure, and the three repeating units are connected in series. The latter output is connected to a convolutional layer. The output of the convolutional layer is spliced with the original input through a convolutional layer. After splicing, batch normalization and activation function Leaky relu connect the CBL convolutional structure at the end.

The characteristics of the CBL convolution structure in FPN are: the size of the convolution kernel is 1 × 1, and the stride is 1.

The characteristics of the CBL convolution structure in PAN are: the size of the convolution kernel is 3 × 3, and the stride is 2.

The SPP module consists of four parallel pooling layers with maximum pooling kernel sizes of 1×1, 5×5, 9×9, and 13×13, respectively. The SPP module itself includes a two-layer CBL convolution structure. part.

The output of the YOLOv5 model inherits the idea of YOLOv3, and uses three scale feature maps for detection, and the feature map sizes are 19×19, 38×38, and 76×76 respectively. And assign a corresponding number of Anchor Boxes to each scale after each SPP module added, generate 9 Anchor Boxes for each pixel in the feature map, filter out the optimal box through weighted non-maximum suppression, and put the GIOU Returned to the network as a loss function, training parameters.

This example is completed under the platform of Centos 7.9.2 and implemented using Python programming. The computer performance of the test network model and the training network model is: Tesla v100, interl Xeon(R) Gold 6271c CPU @2.6GHz; the framework used is pytorch depth Learning framework. The learning rate of the YOLOv5 model is selected as λ=0.01, and the number of training steps is 500 times.

In this embodiment, AP (Average Precision, average precision) and mAP (mean Average Precision, average accuracy) are used for evaluation. In object detection, each category corresponds to a precision rate and a recall rate. The common evaluation indicators in recent years are AP and mAP. AP is the area under the precision-recall curve. The P-R curve is drawn with Precision as the y-axis and Recall as the x-axis. The pros and cons of the evaluation model are mainly determined by the size of the area under the curve. mAP is the average AP of multiple classes.

Among them, TP is the number of positive examples correctly classified into positive examples, FP is the number of positive examples incorrectly classified into negative examples, and FN is the number of negative examples classified into positive examples.

AP is calculated as:

Among them, p represents the precision rate (precision), r represents the recall rate (recall).

In actual statistical situations, precision and recall are not continuous curves, but are independent finite numbers. Therefore, it is discrete statistics during the calculation process:

Among them, N represents the total number of images in the data set to be tested, p(k) represents the accuracy rate of k images recognized by the model, and Δr(k) represents when the model recognizes k images to k-1 images. Changes in recall rate.

In target detection, the degree of overlap between the predicted marked frame and the real marked frame of the target is usually used to judge whether the model correctly detects the target. This degree of overlap is also called IOU (Intersection Over Union), and IoU is generally set. The threshold is 0.5. If the IoU calculated by the model is greater than 0.5, then it is determined that the target is correctly detected. The schematic diagram of the IoU is shown in the figure.

The calculation formula is:

Among them, A∩B represents the overlapping area of the prediction frame and the target frame, and A∪B represents the union area of the prediction frame and the target frame.

In this example, four kinds of defect images, including block defects, scratches, oil spots and white spots on the surface of the strip steel, have been tested. The recognition accuracy of white spots is about 87%, and the recognition rates of all other defects are above 90%. , the recognition rate of block defects and oil spots with similar structures is high, indicating that the detection method of the present invention has high discrimination and high detection accuracy for these two types of similar defects.

What is not described in the present invention applies to the prior art.

Claims (4)

1. A strip steel surface defect detection method based on an improved YOLOv5 network is characterized by comprising the following steps:

the first step is as follows: image dataset acquisition

1.1, acquiring a surface image of the strip steel by using an industrial camera, and screening out a picture containing a defect; when the defect type in the screened defect image covers the known type of the surface defect of the strip steel, a defect image set is formed;

1.2, carrying out size normalization operation on the defect picture set, and then manually labeling the pictures in the defect picture set by using Labelimg software to enable each defect picture to have a label of defect type and defect position coordinates;

1.3 randomly dividing not less than 60% of the marked defect picture set into a training set, and taking the rest as a verification set;

the second step: improved YOLOv5 network model

The improved YOLOv5 network model is characterized in that on the basis of a YOLOv5 network model, a CSA module is connected in series between three CSP23_ modules of a PAN and three conv modules of a classification and positioning part of the network model;

the CSA module comprises a channel attention module and a space attention module, wherein the two modules are connected in series, and the output of the channel attention module is the input of the space attention module;

CSP23_ Module output feature map F₁The input is processed by a channel attention module firstly, and the input characteristic F is processed by the channel attention module firstly₁Respectively carrying out global maximum pooling and global average pooling based on depth and width to obtain two Cx 1 x 1 feature maps; secondly, respectively carrying out fast one-dimensional convolution processing on the two obtained Cx 1 x 1 characteristic graphs with a convolution kernel size of k, adding results obtained by the two fast one-dimensional convolutions, and carrying out sigmoid processing on the results to obtain channel attention; multiplying the channel attention by the original feature F₁Re-weighting the features to obtain weighted features F₂；

Feature F of channel attention module output₂Input to the spatial attention Module, spatial attention Module to feature F₂Respectively carrying out global maximum pooling and global average pooling to obtain two characteristic graphs of 1 multiplied by W multiplied by H; then, performing channel splicing operation on the two 1 xWxH feature graphs based on channels, performing convolution operation with a convolution kernel of 7 x 7 on the result obtained by the splicing operation, reducing the dimension to a one-dimensional vector, namely 1 xWxH, and generating a spatial attention weight through an activation function sigmoid; finally, the spatial attention weight and the input feature F are combined₂Multiplying to obtain the output characteristic F of the space attention module₃(ii) a Characteristic F₃The CSA module is the output of the CSA module and is also the input of the conv module of the classification and positioning part of the network model;

the third step: training improved Yolov5 network model

3.1 image dataset preprocessing

Preprocessing the training set in a Mosaic data enhancement mode;

3.2 parameter settings

Initializing all weight values, bias values and batch normalization scale factor values, setting the initial learning rate and batch _ size of the network, and inputting initialized parameter data into the network; dynamically adjusting the learning rate and the iteration times according to the change of the training loss so as to update the parameters of the whole network; the training is divided into two stages, the first stage is the first 100 periods of the training, and the initial learning rate is fixed to be 0.001 so as to accelerate convergence; the second stage refers to a training period 100 periods later, and the initial learning rate is set to 0.0001;

3.3 network model training

Inputting the preprocessed training set into an improved YOLOv5 network model with set initialization parameters in the second step for feature extraction, automatically generating an anchor frame for the images of the training set by using a K-means clustering method, taking the size of the anchor frame as a prior frame, and obtaining a boundary frame through frame regression prediction; then, classifying the bounding boxes by using a logistic classifier to obtain defect class classification probability corresponding to each bounding box; sorting the defect type classification probabilities of all the boundary frames by a non-maximum value inhibition method, and determining the defect type corresponding to each boundary frame to obtain a predicted value; the predicted value comprises defect type and defect position information, and the non-maximum inhibition threshold value is 0.5; then calculating a loss value between the predicted value and the true value through a loss function GIOU; performing back propagation according to the training loss value, updating parameters of the backbone network and the classification regression network until the loss value accords with the preset value, and finishing the training of the network model parameters;

3.4 network model testing

Inputting the verification set into the network model which completes parameter training in the step 3.3 to obtain a tensor prediction value of the verification set; comparing the predicted value of the tensor with the labeling information, and testing the reliability of the network model; evaluating the network model by using the AP, and testing the network model to be reliable when the AP is not less than 85%;

the fourth step: strip steel surface defect detection

And (3) performing the same size normalization operation as in the step 1.2 of the first step on the surface image of the strip steel to be detected, and then inputting the image into the network model tested to be reliable in the third step to obtain the defect tensor information of the surface image of the strip steel to be detected, including the defect position, the defect type and the confidence coefficient.

2. The strip steel surface defect detection method based on the improved YOLOv5 network as claimed in claim 1, wherein the convolution kernel size k of the fast one-dimensional convolution in the channel attention module represents the coverage of local cross-channel interaction, i.e. how many neighbors participate in the attention prediction of the one-dimensional channel; wherein the coverage of the interaction k is proportional to the channel dimension, and the specific calculation formula is:

where C denotes the number of eigenchannels, β ═ 2 and b ═ 1 denote two superparameters.

3. The method as claimed in claim 1, wherein in step 1.2 of the first step, the size normalization is performed to scale the image to 608 × 608 pixels.

4. The method for detecting the surface defects of the steel strip based on the improved YOLOv5 network as claimed in claim 1, wherein in step 1.3 of the first step, the training set is 80%, and the rest 20% is the verification set.

Images