OPTIMIZED YOLOV5 FOR SOLAR CELL SURFACE DEFECT DETECTION

A solar cell defect detection method with an 
improved YOLO v5 algorithm is proposed 
for the characteristics of the complex solar 
cell image background, variable defect 
morphology, and large-scale differences. 
First, the deformable convolution is 
incorporated into the CSP module to achieve 
an adaptive learning scale and perceptual 
field size; then, the feature extraction 
capability of the model is enhanced by 
introducing the ECA-Net attention 
mechanism; finally, the model network 
structure is improved and one tiny defect 
prediction head is added to improve the 
accuracy of target detection at different 
scales.