HF-YOLO:基于特征协同的低照度钢材缺陷检测OA

In the task of detecting steel surface defects under low illumination conditions,due to the weak saliency of target features and the difficulty in distinguishing between targets and backgrounds,missed detections are prone to occur.Therefore,a feature collaborative low illumination steel surface defect detection method,named HF-YOLO,was proposed.Firstly,a high and low frequency feature extraction module(HLFM)was designed to process high frequency and low frequency features of images,extract target and edge information.Secondly,a dual branch cross path collaboration module(DBCM)was designed to extract rich contextual information in both shallow and deep layers.The feature reshaping and mapping module(FRMM)was utilized to reshape and perform linear mapping on features,establishing spatial and semantic associations between information.Finally,transfer learning was introduced by training the model on the source domain low-light dataset ExDark to learn how to process low-light image features,and the training parameters are transferred to the low light steel surface defect detection task to improve the detection performance of the model.Additionally,a steel surface defect dataset named Dark-SSD,encompassing varying degrees of low-light conditions,was constructed to provide data support for the low-light steel surface defect detection task.Experimental results show that HF-YOLO achieves a mAP of 77.2%on the NEU-DET dataset,representing a 3.2%improvement compared to the baseline model.On the self-built low illumination dataset Dark-SSD,the mAP is improved by 4.3%.The experimental results demonstrate the effectiveness and feasibility of the HF-YOLO method,which is suitable for the task of detecting steel surface defects under low illumination.