Deep Learning-based Industrial Product Defect Detection Algorithm

In today’s society,industrial products are omnipresent.Due to a variety of factors,these products inevitably develop defects during their manufacturing process,which can adversely affect their quality and performance.Therefore,ensuring product quality and maintaining stable production have made industrial defect detection critically important.Traditional manual quality inspection is not only costly and inefficient,but also struggles to meet the demands of large-scale quality control.In recent years,the rapid advancements in technologies such as industrial imaging,computer vision,and deep learning have made their application to industrial defect detection particularly significant.However,utilizing supervised learning for defect detection presents challenges such as a scarcity of defect samples,irregular defect shapes,and unknown defect types.Acquiring a diverse range of defect samples is difficult,leading to potential missed detections or false positives.This paper proposes a deep learning-based method for industrial defect detection,which is capable of achieving high-precision detection requirements by solely training models with normal images.This approach can effectively manage product quality,reduce inspection costs,and broaden the scope of inspection.For the field of industrial production,such a defect detection method is indispensable.The research content of this paper includes the following points:(1)We propose an unsupervised industrial anomaly detection algorithm based on the block pyramid memory module.This module is designed to store data from positive samples and is capable of filtering out anomalous features by accessing data from positive samples across different scales.This process significantly improves the suppression of larger anomalies during reconstruction and enhances the precision of the reconstructed details.We use skip connections to deliver feature maps — enhanced by the block pyramid memory module at different scales— into the decoder,thereby increasing the clarity of the reconstructed images.Additionally,a patch-based SSIM loss function is utilized to amplify differences between the original and reconstructed images.(2)We propose a self-supervised industrial anomaly detection algorithm based on the generation of fake anomaly samples.This algorithm trains the model using only normal images and fake defect images,generated through the salient anomaly fake method based on Perlin noise and Bézier curves(SAFM-PB),thereby increasing the diversity and irregularity of the fake defects.We propose a feature fusion module based on spatial attention mechanism(FFMSA),effectively merging features extracted by the encoder and multi-scale features filtered of anomalies by the block pyramid memory module,to enhance the detail information of the reconstructed image and improve the accuracy of anomaly detection.Finally,the multi-scale gradient magnitude similarity(MSGMS)function is utilized to compare the structural differences between the input and reconstructed images,enhancing the detection performance.(3)We propose an industrial anomaly segmentation algorithm based on self-supervised learning.This algorithm learns the decision boundaries between normal and abnormal samples by learning the joint representation of anomaly images and reconstructed images,thereby directly performing anomaly localization and avoiding the need for complex post-processing steps.Firstly,multi-scale features are extracted from the anomaly and reconstructed images,with the cascade of features at corresponding scales providing more effective information for anomaly region localization.The algorithm addresses the issue of multi-scale industrial defects by employing multi-scale feature fusion.Furthermore,through the addition of the Coordinate Attention(CA)mechanism,the algorithm captures cross-channel information,orientation perception,and location information,assisting the model in more accurately locating and identifying objects of interest.The method in this paper performs excellently on the standard dataset MVTec AD.The average AUROC score for image-level anomaly detection reaches 99.1%,and the average AUROC score for pixel-level anomaly detection reaches 97.0%,surpassing many of the most advanced anomaly detection algorithms currently available.