CN114897857A - Solar cell defect detection method based on light neural network - Google Patents

Abstract

The invention relates to a solar cell defect detection method based on a light weight neural network, wherein a surface defect detection model constructed by the method is based on a light weight attention YOLOv5 network and consists of an input end, a backbone network, a characteristic enhancement part and an output end; the input end is used for data enhancement, the backbone network is composed of a Shufflenetv2 network and a global context information module, the Shufflenetv2 network firstly carries out convolution, batch normalization, ReLU activation function and pooling operation on the feature map for one time, then the feature map alternately enters a stage module composed of a spatial down-sampling unit and a basic unit in sequence, and the operation is repeated for three times to obtain a final output feature map; the characteristic enhancement part consists of a characteristic pyramid and a path enhancement network, and further fusion enhancement is carried out on information by adopting paths from bottom to top and from top to bottom; the output is used to calculate the training loss of the model. The model has less parameter quantity, reduces the calculation complexity of the model, and can be deployed at a mobile terminal.

Description

Solar cell defect detection method based on light neural network

Technical Field

The invention relates to the technical field of image defect detection, in particular to a solar cell defect detection method based on a light weight neural network.

Background

A solar cell is a cell that converts solar radiation energy directly or indirectly into electrical energy through a photoelectric effect or a photochemical effect by absorbing sunlight. In the production and transportation process of the solar cell, the cell is inevitably lost, so that surface defects are caused, and common defect types mainly comprise thick lines, broken grids, slurry leakage, color difference, dirty sheets, scratches and the like. Because the batteries are mostly mounted in an array mode, the performance of the whole battery can be reduced due to the defect of a small battery, and therefore, the detection of the defect on the surface of the solar battery plays an important role in the production process. Dense grid lines are arranged on the solar cell, forming a regular texture on the surface, which makes surface defect detection very difficult. At present, most solar cell production plants mainly adopt a manual detection mode, visual judgment of workers is relied on, the manual detection mode is easy to cause eye fatigue, and the manual detection mode is influenced by artificial subjective factors to cause misjudgment.

With the development of computer vision technology and the improvement of computing power, the computer vision technology is widely applied to various industries, and the computer vision technology not only can overcome the defects of a manual detection mode, but also can improve the production efficiency and the quality rate of products.

The patent application with the application number of 202110646252.1 discloses a solar cell surface defect detection method, which comprises the steps of further extracting and fusing the characteristics of different scales and different receptive fields through a three-branch cavity rolling block, and then classifying and regressing through an RPN classification and regression branch network respectively, but the method only has good detection effects on two defects, namely a thick line and a crack, and has poor detection effects on rosin joint; meanwhile, the amount of parameters of a backbone network model based on Resnet101 is too large, which is not beneficial to actual deployment. The patent application No. 202010429805.3 introduces the idea of cross-layer connection on the basis of the Faster R-CNN convolutional neural network, and extracts the target candidate box in a multi-scale manner, but because the method adopts the Faster R-CNN convolutional neural network, the parameter amount and the calculation amount are very large and the real-time performance cannot be guaranteed because the method is a two-stage network.

In order to achieve a higher detection effect, the network structure of the CNN is also becoming more and more complex, the amount of parameters is increasing, and the computational complexity is gradually increasing. For embedded devices which cannot meet the real-time requirement, a lightweight solar cell defect detection model needs to be developed urgently.

Disclosure of Invention

Aiming at the defects in the prior art, the invention aims to provide a solar cell defect detection method based on a lightweight neural network.

The technical scheme adopted by the invention for solving the technical problems is as follows:

a solar cell defect detection method based on a lightweight neural network is characterized by comprising the following steps:

the first step is as follows: image pre-processing

1) Establishing a solar cell defect image library: firstly, acquiring a solar cell defect image by using an industrial camera, then performing operations such as rotation, translation, contrast deepening and the like on the defect image to obtain a preprocessed image, and finally performing size normalization operation on the defect image and the preprocessed image to obtain images with consistent sizes, and establishing the images as a solar cell defect image library;

2) data set preparation: and selecting a training set, a verification set and a test set in a random selection mode from the solar cell defect image library. Marking each image in the training set, the verification set and the test set respectively, and adding defect type and defect position information;

secondly, building a surface defect detection model;

the surface defect detection model is based on a light attention YOLOv5 network and consists of an input end, a backbone network, a characteristic enhancement part and an output end;

the backbone network is composed of a Shufflentv 2 network and a global context information module, the Shufflentv 2 network firstly carries out convolution, batch normalization, ReLU activation function and pooling operation on the feature map once, then sequentially and alternately enters a stage module consisting of a spatial down-sampling unit and a basic unit, and the operation is repeated for three times to obtain a final output feature map;

after the channels of each spatial down-sampling unit and the basic unit are shuffled, a global context information module is connected, the global context information module copies the input feature maps into two groups, one group passes through a context modeling module and a conversion module in sequence, the other group does not operate, and finally the two groups are added; the global context information module comprises a context modeling module and a conversion module, wherein the context modeling module copies the input feature maps into two groups, one group of the feature maps is subjected to 1 × 1 convolution and softmax activation function operation in sequence, weights are taken from the input feature maps, the other group of the feature maps is not subjected to operation, and finally the two groups of feature maps are multiplied to carry out weight assignment on the input feature maps so as to extract context information; the conversion module sequentially performs 1 × 1 convolution, layer normalization, a ReLU activation function and 1 × 1 convolution operations on the input feature graph, and is used for extracting features and adjusting the size of the feature graph; the characteristic enhancement part consists of a characteristic pyramid and a path enhancement network, and further fusion enhancement is carried out on the information by adopting paths from bottom to top and from top to bottom.

And thirdly, training the surface defect detection model, and using the trained surface defect detection model for detecting the defects of the solar cell.

Compared with the prior art, the invention has the beneficial effects that:

1. aiming at the problems that a target detection model based on a deep convolutional neural network has large parameter quantity and can not meet the real-time requirement for some embedded equipment with limited resources, the Shufflentv 2 is selected as a backbone network, so that the parameter quantity of the model is reduced, the calculation complexity of the model is reduced, and the detection model can be deployed at a mobile terminal; a global context information module is introduced into the Shufflentv 2 network so as to overcome the problem that the detection accuracy of the Shufflentv 2 network is low in the training process.

2. The introduced global context information module can inhibit background information of the defect detection model in the training process, highlight foreground information and reduce the influence of model precision loss caused by the reduction of backbone network parameters.

Drawings

FIG. 1 is a schematic diagram of the overall structure of a surface defect inspection model according to the present invention;

fig. 2 is a schematic structural diagram of a basic unit in a backbone network;

FIG. 3 is a schematic diagram of a spatial down-sampling unit in the backbone network;

fig. 4 is a schematic structural diagram of the global context information module.

Detailed Description

The technical solutions of the present invention are described in detail below with reference to the accompanying drawings and specific embodiments, but the scope of the present invention is not limited thereto.

The invention provides a solar cell defect detection method (short method) based on a light neural network, which comprises the following steps:

firstly, preprocessing an image;

1) establishing a solar cell defect image library: firstly, acquiring a solar cell defect image by using an industrial camera, then performing operations such as rotation, translation, contrast deepening and the like on the defect image to obtain a preprocessed image, and finally performing size normalization operation on the defect image and the preprocessed image to obtain images with consistent sizes, and establishing the images as a solar cell defect image library; the size of an original picture used in this embodiment is 649 × 649, the size of the picture sent to training is 416 × 416, and 6396 polycrystalline silicon solar cell defect image libraries are obtained after image preprocessing;

2) data set preparation: and selecting a training set, a verification set and a test set in a random selection mode from the solar cell defect image library, wherein the ratio of the training set to the test set is 8:1: 1. Marking each image in the training set, the verification set and the test set respectively, and adding defect type and defect position information; the marked object is an image defect area, the defect types comprise thick lines, broken grids, scratches, slurry leakage, color difference and dirty films, and the defect area is marked manually by using Labelimg; in this embodiment, 5116, 639, and 641 defect pictures are selected from the training set, the verification set, and the test set, respectively;

secondly, building a surface defect detection model;

the surface defect detection model is based on a light-weight attention YOLOv5 network, and the specific structure is shown in FIG. 1, and mainly comprises an input end, a backbone network, a feature enhancement part and an output end;

the input end mainly performs mosaic data enhancement on the image, and the input image is spliced in a random zooming, random cutting and random arrangement mode, so that the detection effect on the small target is improved;

the backbone network is obtained by combining an improved Shufflentv 2 network with a global context information module (GCBlock), and mainly provides characteristics for an input characteristic diagram; the input feature map passes through a Shufflentv 2 network and a global context information module to extract image features; the Shufflentv 2 network firstly carries out convolution, batch normalization, ReLU activation function and pooling operation on the feature map once, then sequentially and alternately enters a stage module consisting of a spatial down-sampling unit and a basic unit, and repeats the operation three times to obtain a final output feature map; in particular, the convolution, batch normalization, ReLU activation function, pooling, spatial down-sampling unit and base unit repetition times, and output channel output in this embodiment are shown in table 1.

A global context information module (GCBlock) is connected behind the Channel shuffle (Channel shuffle) of each spatial down-sampling unit and the basic unit, the global context information module copies the input feature maps into two groups, one group passes through a context modeling module and a conversion module in sequence, the other group does not operate, and finally the two groups are added; the global context information module comprises a context modeling module and a conversion module, wherein the context modeling module copies the input feature maps into two groups, one group of the feature maps is subjected to 1 × 1 convolution and softmax activation function operation in sequence, weights are taken from the input feature maps, the other group of the feature maps is not subjected to operation, and finally the two groups of feature maps are multiplied to carry out weight assignment on the input feature maps so as to extract context information; the conversion module sequentially performs 1 × 1 convolution, layer normalization, a ReLU activation function and 1 × 1 convolution operations on the input feature graph, and is used for extracting features and adjusting the size of the feature graph;

TABLE 1 backbone network integral composition structure

The characteristic enhancement part consists of a characteristic pyramid and a path enhancement network, and further fuses and enhances information by adopting paths from bottom to top and from top to bottom.

The output end uses GIOU _ Loss as a Loss function of the boundary box, so that the problem of non-coincidence of the boundary boxes is effectively solved, and the regression speed and precision of the prediction box are improved;

thirdly, training a surface defect detection model, and using the trained surface defect detection model for detecting defects of the solar cell;

1) preprocessing a training set: preprocessing the training set in a mosaics data enhancement mode; the Mosaic data enhancement adopts the modes of random zooming, random cutting and random arrangement to splice the images, so that the small target detection effect is good;

2) setting parameters: initializing all training parameters, setting the maximum iteration number to be 200, setting the initial learning rate to be 0.01, setting the batch size (batch _ size) to be 32, and setting the weight attenuation coefficient to be 0.001;

3) training a network model: inputting the preprocessed training set into a surface defect detection model with all initialized training parameters for training; firstly, extracting features by using a backbone network, then entering a feature enhancement part consisting of a feature pyramid and a path enhancement network, further fusing information by adopting paths from bottom to top and from top to bottom, and finally predicting the position, the type and the confidence coefficient of the data defects of a training set through an output end; comparing the generated predicted value with the marking information to generate a loss value, and performing reverse propagation updating until the loss value is not changed any more, and finishing the training of the network model;

4) testing a network model: inputting the verification set into a trained surface defect detection model to obtain a predicted value of data of the verification set, comparing the predicted value with the labeling information in the first step, checking the reliability of the surface defect detection model, and checking whether the surface defect detection model has the problems of over-fitting and under-fitting so as to determine whether the training needs to be terminated and parameters need to be adjusted again;

inputting the test set into a trained surface defect detection model to obtain prediction information of a defect image of the test set, wherein the prediction information comprises defect position information, category information and confidence coefficient information;

therefore, the defect detection of the solar cell based on the light weight neural network is completed.

In order to verify the effectiveness of the method of the present invention, the surface defect detection model of the present invention and the YOLOv5s network were used to detect solar cell defects, respectively, and the comparative test results shown in table 2 were obtained.

Table 2 comparative test results

MAP0.5 and MAP0.5-0.95 are used as evaluation indexes, wherein MAP0.5 refers to the average precision of a network model for a mean value of a detection data set when the intersection ratio (IOU) of a prediction frame and a real frame is more than or equal to 0.5; MAP0.5-0.95 refers to the mean average accuracy at the intersection ratio of prediction and real boxes (IOU) at (0.5, 0.55, 0.6, 0.65, 0.7, 0.75, 0.8, 0.85, 0.9, 0.95). As can be seen from table 2, when the detection accuracy is sufficiently high, the detection model based on S2-YOLOv5 has smaller parameter number and computational complexity, and can realize fast and light detection of defects of the solar cell.

Detecting the defects of the solar cells of different types by using the surface defect detection model disclosed by the invention to obtain the test results shown in the table 3; the surface defect detection model disclosed by the invention has higher recognition rate on six types of defects, and can realize accurate detection of the defects of the solar cell.

TABLE 3 test results for different defect types

Kind of defect	Thick line	Broken grid	Scratch mark	Slurry leakage	Color difference	Dirty sheet
							MAP0.5(％)	94.7	94.6	83.2	95.6	99.4	87.7
MAP0.5-0.95(％)	57.8	44.9	48.3	56.4	87.2	51.5

Nothing in this specification is said to apply to the prior art.

Claims (1)

1. A solar cell defect detection method based on a lightweight neural network is characterized by comprising the following steps:

the first step is as follows: image pre-processing

1) Establishing a solar cell defect image library: firstly, acquiring a solar cell defect image by using an industrial camera, then performing operations such as rotation, translation, contrast deepening and the like on the defect image to obtain a preprocessed image, and finally performing size normalization operation on the defect image and the preprocessed image to obtain images with consistent sizes, and establishing the images as a solar cell defect image library;

2) data set preparation: and selecting a training set, a verification set and a test set in a random selection mode from the solar cell defect image library. Marking each image in the training set, the verification set and the test set respectively, and adding defect type and defect position information;

secondly, building a surface defect detection model;

the surface defect detection model is based on a light attention YOLOv5 network and consists of an input end, a backbone network, a characteristic enhancement part and an output end;

the backbone network is composed of a Shufflentv 2 network and a global context information module, the Shufflentv 2 network firstly carries out convolution, batch normalization, ReLU activation function and pooling operation on the feature map once, then sequentially and alternately enters a stage module consisting of a spatial down-sampling unit and a basic unit, and the operation is repeated for three times to obtain a final output feature map;

after the channels of each spatial down-sampling unit and the basic unit are shuffled, a global context information module is connected, the global context information module copies the input feature maps into two groups, one group passes through a context modeling module and a conversion module in sequence, the other group does not operate, and finally the two groups are added; the global context information module comprises a context modeling module and a conversion module, the context modeling module copies the input feature maps into two groups, one group of the feature maps is sequentially subjected to 1 multiplied by 1 convolution and softmax activation function operation, the weight of the input feature map is taken, the other group of the feature maps is not operated, and finally the two groups of the feature maps are multiplied to carry out weight assignment on the input feature map, so that the context information is extracted; the conversion module sequentially performs 1 × 1 convolution, layer normalization, a ReLU activation function and 1 × 1 convolution operations on the input feature graph, and is used for extracting features and adjusting the size of the feature graph; the characteristic enhancement part consists of a characteristic pyramid and a path enhancement network, and further fusion enhancement is carried out on the information by adopting paths from bottom to top and from top to bottom.

And thirdly, training the surface defect detection model, and using the trained surface defect detection model for detecting the defects of the solar cell.

Images