基于可改变核卷积的不规则织物疵点识别算法研究OA

An improved YOLOv7 feature extraction network was proposed to solve the problems of fabric defects in textile industry,such as variable shapes and small target ranges.The variable nuclear convolution is introduced into the YOLOv7 feature network to replace the traditional convolution block,and the arbitrary sampling shape and arbitrary parameter number characteristics of the variable nuclear convolution are used to improve the efficiency of collecting defect feature information,which can better adapt to irregular defect shape and size characteristics during sampling.At the same time,efficient multi-scale attention is embedded,and interdimensional interactions are used to capture pixel-level relationships and improve the ability of the model to process tiny features.Through experimental verification,the detection accuracy rate P of various samples of the improved model reaches 95.1%,which is 7.3%higher than that of YOLOv7 baseline model,the recall rate R is 7.2%higher,and mAP@0.5 is 12.3%higher.The detection accuracy and detection speed are greatly improved.The improved model is more efficient to identify irregular fabric defects and objects with smaller defect range,and can provide technical support to efficient and rapid identification of fabric defects in industrial scenes.