Attention-Based Surface Defect Weakly-Supervised Detection Of Polycrystalline Silicon Solar Cells

Surface defect detection has become one of the key processes in the production process of photovoltaic cells.The types silicon photovoltaic cells of surface defects include paste spot,broken grids,scratches,color difference,thick lines,and dirty cells.The backgrounds,colors,textures,and shapes of defective cells vary from each other.At the same time,due to the low incidence of defects and complex process of manually labeling,the cost of collecting accurately labeled samples for model training is extremely expensive,which also limits the application of deep learning methods’ application in the field of surface defect detection.In the context of intelligent manufacturing,this paper proposed an inspection method based on weakly supervised learning.The framework improved the existing deep learningbased visual detection framework for surface defects,reduced the dependence on accurately labeled defect samples,and improved the effect of classification and segmentation for solar cell surface defects.Specifically,by optimizing the structure of the existing convolutional neural network model,a novel model with attention mechanism and min-maximum entropy was proposed,which significantly improved the ability of detect classification and feature extracting.Then,the improved class activation mapping model is used to further improve the effect of photovoltaic cell surface defect detection.At the same time,the proposed weakly supervised machine learning method does not require precise pixel-level labeling to achieve accurate defect region segmentation effects,which reduced the cost of manual labeling samples.The contributions of this article are as follows:1.In order to solve the problem of multi-class defect classification under the photovoltaic cells background of non-uniform complex texture,this paper used the existing convolutional neural network structure and deep learning dataset of photovoltaic cell surface defects collected from the production line.The different network structures were compared and best of them was selected for the segmentation task.Further,the classification layer of the convolutional neural network was replaced by a random forest classifier,which made CNN more robust for defect classification tasks under complex backgrounds such as polycrystalline silicon photovoltaic cells.2.In order to improve the segmentation effect,this paper proposed a defect inspecting framework(CBA-CAM)based on densely connected attention networks,which significantly improved the accuracy and robustness of the classification effect.Furthermore,by introducing attention mechanism in the side connection of densely connected networks,suppressing the background,highlighting the response of defect areas,and selecting corresponding stronger features,the model’s ability to distinguish between texture background and defect features was enhanced.In the experiment,K-fold cross validation was used to traverse all experimental pictures to prove that the model was not overfitted.The final experimental results showed that the proposed intensive attention-connected defect detection framework could effectively deal with the complex surface background of solar cell,accurately classify defects with diverse features and random shapes.Based on this,the model was generate to different datasets.3.Based on the above model,to further improve the weakly supervised defect segmentation effect,this paper proposes a class activation mapping model based on tandem attention and min-maximum entropy(MME-ACAM).On the basis of densely connected attention,a new optimization index during training—min-maximum entropy is introduced instead of the traditional cross-entropy.This method further improves the robustness and segmentation effect of the weakly supervised method of CAM-based for the surface defect detection and segmentation problems of polycrystalline silicon batteries.Under the action of min-maximum entropy and attention mechanism,the improved MME-ACAM model discriminated more accurate background and defection features.The segmentation effect had been significantly improved.