Research On Crack Detection Of Switch Panel Based On CNN

Cracks in products seriously affect the quality of the products.Timely detection of cracks in products can effectively improve product quality and increase consumer satisfaction with the products.This article focuses on the detection of cracks on the surface of switches and explores the effective detection of crack features using computer vision technology.This article mainly focuses on the surface crack detection of switches,and is based on the research of using convolutional neural networks to detect the characteristics of surface cracks on switches.The main research contents are as follows:1.This study investigates the surface crack detection of switches and captures images of surface cracks on switches,creating a dataset of switch surface images that includes crack images.The cracks were classified based on their length,with those less than 0.5 centimeters categorized as fine cracks,those with lengths between 0.5-2.0 centimeters categorized as medium cracks,and those with lengths greater than 2.0 centimeters categorized as large cracks..2.To address the issue of difficulty in obtaining crack images and the problem of irregular image collection,the cracks were preprocessed and data augmentation was performed.In addition,knowledge distillation and transfer learning were employed during network training to accelerate network convergence.3.Aiming at the problem of crack defect identification,the Rep VGG Block based on structural re-parameterization is integrated into the efficient channel attention network and Gaussian linear error unit,and the method of distillation learning is used to train during training.Finally,the KD-EG-Rep VGG network is obtained.In crack detection,the accuracy of the network is more than 98%,the calculation amount of the model is only 0.03 G,and the FPS is more than400frame/s.4.To address the issue of detecting crack defects,we have enhanced the twostage Faster R-CNN network.Specifically,we have incorporated the KD-EGRep VGG22 network as the feature extraction network,utilized the ROI Align algorithm to improve the ROI pooling process,and added a channel pooling layer to the full connection layer in order to reduce the model’s parameter count.Through our experiments,we have observed that these improvements have led to a crack detection accuracy of 93.4% and a reasoning speed of 20 FPS,which fulfills the real-time requirements.