高频引导的多尺度空间感知异常检测网络OA

Aiming at the problems of low detection accuracy for small-sized defects,weak multi-scale fea-ture extraction ability and low anomaly segmentation accuracy of existing industrial anomaly detection algo-rithms,an industrial anomaly detection network combining high-frequency residual guidance and multi-scale attention feature fusion was proposed.Firstly,aiming at the problem of high-frequency detail loss caused by traditional full-frequency processing,a frequency-domain separation strategy was designed.Gaussian kernel filtering was utilized to extract high-frequency residual features,enhancing the network's detection ability for minor anomalies.Secondly,aiming at the problems of insufficient representation abili-ty of complex textures and low discrimination between anomalies and backgrounds in conventional convolu-tional networks,a globally enhanced multi-scale attention module GEMA is embedded in the encoder stage of the discriminative network.It captures multi-scale local information in the horizontal and vertical directions through parallel dual-path,enhancing the salient features at different spatial positions.Improve the feature discriminability in complex texture backgrounds;Finally,in the decoder stage of the discrimi-nant network,the coordinate attention module CoordAtt is integrated.By decomposing the coordinate ax-es and dynamically modulating the feature weights,precise spatial positioning of abnormal areas is achieved.Experiments show that on the MVTec AD public dataset,the average AUROC at the image level of the improved model is 98.6％,and the average AUROC and AP at the pixel level are 97.6％and 73.2％respectively,effectively improving the effect of industrial anomaly detection.