改进Faster RCNN with FPN的素布瑕疵检测的算法研究OA

In the textile industry,manual detection is still used for cloth inspection.The effect of manual detec-tion is greatly affected by workers'subjectivity,leading to reduced efficiency and missed or erroneous detection of defects.To address this situation,the algorithm of plain cloth defect detection is explored to improve Faster RCNN with FPN target detection algorithm.Firstly,in order to improve the fusion capability of Faster RCNN with FPN for multi-scale features and enrich the context information of each feature layer,the cross-scale fea-ture fusion module is introduced to improve the feature pyramid network structure.Secondly,in order to make better use of deep features,the intra-scale feature interaction module is added to process the deep feature layer output by ResNet50,and enrich the semantic information of the high-level feature layer.Then,in order to en-hance the detection capability for defects of extreme sizes,preset anchor boxes are improved by using K-means++ clustering and genetic algorithms.Finally,considering the small size of plain fabric defects,Focal Loss is used to balance the positive and negative samples to increase the detection effect of plain cloth defects.Through experiments,the COCO index is used for evaluation.Compared with Faster RCNN with FPN,the im-proved network model increases by 6.5%,4.4%and 4.0%on the mAP50,mAP75 and mAP50:95 indexes,respective-ly.The average accuracy has been significantly improved,which can better complete the detection task of plain cloth defects.