Research On Anomaly Detection And Location Algorithm Of Industrial Image Based On Deep Learning

With the deep integration of the traditional manufacturing industry and the new generation of information technology,industrial products have gradually shifted from quantity expansion to quality improvement,thus effectively improving the high added value and competitiveness of industrial products.The detection of product surface quality is a significant means to ensure product quality,mainly including traditional manual quality inspection,quality detection based on machine vision,deep learning and other methods.At present,most of the surface anomaly detection methods of industrial products have some problems,such as low detection accuracy,poor generalization performance,low accuracy,poor real-time performance and low efficiency.This paper designs two different methods of anomaly detection and location in industrial images.The primary research contents and accomplishments are as follows.(1)Most current anomaly detection methods mainly use normal samples or unmarked data for training.Due to the lack of prior anomaly knowledge,it is easy to misjudge normal samples with noise data as anomalies.This paper proposes a weak supervised anomaly detection model based on deep anomaly scoring network.In this model,ResidualNetwork(ResNet)is used as feature extraction network,and Res2Net module is added to the ResNet,which represents multi-scale features in a more fine-grained level to improve the multi-scale representation ability of the network.At the same time,Efficient Channel Attention(ECA)is introduced to enhance feature extraction performance by allocating the attention of feature channels.The anomaly score network calculates the anomaly score directly according to the extracted feature representation,and optimizes the anomaly score in an end-to-end manner.The comprehensive experiments on MVTec AD,KolektorSDD and ELPV datasets show that the model achieves better results and robustness in anomaly detection and location compared with the current mainstream anomaly detection methods.(2)In view of the shortcomings of most current industrial product surface defect detection methods,such as low detection accuracy and poor generalization performance,this paper proposes a reconstructive and discriminative anomaly detection model(RDAD).The model uses Sequence-and-Excitation block(SE block)to assign the attention of feature channels to enhance the sensitivity of related features and improve the ability of the model to learn normal and anomaly boundaries.In addition,the Channel Transformer(CTrans)is introduced in the encoder-decoder,so that the decoder better fuses the features in the encoder and reduces the semantic gap,and enhances the segmentation ability of anomalous regions of the model.The model is only trained with normal samples and completes the localization of anomalous regions while detecting anomalies.Experiments are conducted on the challenging MVTec AD and Magnetic Tile Defect datasets.Compared with the current state-of-the-art unsupervised anomaly detection methods,the model not only improves the accuracy of anomaly detection,but also has better generality.