Research On Industrial Defect Detection Algorithm Based On Anchor-free Detector And Knowledge Distillation

With the continuous development of computer vision technology,deep learning methods based on convolutional neural network(CNN)and Transformer have rapidly developed in the field of industrial defect detection,gradually replacing the traditional,time-consuming and laborious manual detection methods.However,there are still some pain points in this field.On the one hand,industrial defects are more complex than natural image targets in terms of background and have a large number of defects with large aspect ratios.The target detection models suitable for natural images will have a problem of accuracy degradation in industrial defect detection.On the other hand,as the model accuracy continues to improve,the computational overhead required also continues to grow,while the hardware equipment used by factories is usually low-performance and cannot afford the computational overhead brought by the large number of parameters on large models,and the real-time performance of detection is poor.Therefore,how to improve the detection accuracy of the model while reducing the model size as much as possible has become a research hotspot in the field of industrial defect detection.In order to better achieve industrial defect detection,this paper proposes a knowledge distillation method based on anchor-free detector D-CenterNet(Distillation-CenterNet),whose main work and innovation are as follows:1)Our method is based on the anchor-free detection algorithm-CenterNet.In view of the problem that the CenterNet algorithm lacks the center point information of large aspect ratio defects during the training process,which leads to a decrease in accuracy,we propose an adaptive label encoding strategy,which can better regress the center point information of defects with large aspect ratios and improve detection performance,reduce background interference and negative sample generation.In view of the characteristics of industrial defect images with more complex background and a large number of defects with large aspect ratios,we propose ResNet-SP network based on ResNet,and introduce Multi-scale strip pooling module(MSP)and branch pooling module(BPM)in ResNet.The multi-scale strip pooling module increases the receptive field by introducing dilated convolution,and uses rectangular pooling kernels instead of basic square pooling kernels,which can more effectively extract the long-distance dependency relationship of defects with large aspect ratios and retain their most significant features.The branch pooling module adopts a parallel structure,uses strip pooling to model the long-distance relationship of defects with large aspect ratios,and uses improved dilated spatial pyramid pooling to extract the local context information of regular defects,avoiding information loss,and finally outputs the fused feature information.The experimental results based on the publicly available Alibaba Cloud Tianchi fabric defect dataset show that compared with the basic CenterNet algorithm and other industrial defect detection algorithms,our proposed algorithm has a significant improvement in accuracy.2)Different from the traditional knowledge distillation method based on model feature layer information,we regard the prediction information output by the model detection head as knowledge that can be used to guide the training of the student model,and propose a knowledge distillation method based on the detection head output information and feature layer foreground information.In view of the problem of too much redundant noise information in the feature layer,we use adaptive Gaussian mask to model the foreground information in the feature map,highlighting the foreground target information while suppressing the redundant information in the background.By transferring the feature knowledge of teacher model’s feature layer and detection head output to lightweight student model through knowledge distillation,using teacher model’s pseudo-labels and real labels in dataset to jointly supervise student model’s training in offline distillation mode,we can improve student model’s performance without loss without increasing its parameter amount.The experimental results based on publicly available Alibaba Cloud Tianchi fabric defect dataset and Northeastern University hot-rolled steel surface defect detection dataset(NEU-DET)show that our proposed model compression method based on knowledge distillation can improve student model’s m AP from 60.1% to 63.8% without increasing model complexity,greatly improving detection accuracy while keeping model size and detection speed basically unchanged.