基于上下文信息聚合YOLOv5的织物缺陷检测OACSTPCD

In response to the problem that the current mainstream inspection methods cannot effectively capture the rich contextual information of fabric defects,resulting in lower detection accuracy and higher leakage rate of defects similar with background,an improved YOLOv5 model(YOLO-CA model)that could aggregate contextual information was proposed.Defects with context-rich contextual relationships were captured by the YOLO-CA model by incorporating a dual cross-attention module(DCCA)in the backbone network of YOLOv5,which enabled the model to focus on the defective area and improved the detection rate of defects by the model.More detailed defect feature information was extracted in combination with the CSP-SoftSPPF module to improve the detection accuracy of the model.The experimental results showed that the YOLO-CA model has mAP values of 67.6%and 93.2%,as well as MR values of 38.4%and 10.5%on SDCFD and TILDA publicly available datasets.Compared with original YOLOv5 model,the mAP value of YOLO-CA model was increased by 20.8 percentage points and 4.5 percentage points,and the MR values were decreased by 16.5 percentage points and 6.9 percentage points respectively.It is considered that the YOLO-CA model has better detection performance,can effectively reduce the leakage rate of defects of similar classes to the background,improves the detection accuracy of several defects,and is suitable for fabric defect detection tasks.