【作者】 刘坤； 刘铁旭； 陈海永； 陈鹏；

【Author】 Liu Kun;Liu Tie-xu;Chen Hai-yong;Chen Peng;College of Artificial Intelligence,Hebei University of Technology;

【机构】 河北工业大学人工智能与数据科学学院；

【摘要】 带钢表面缺陷复杂多变、缺陷尺度跨度大、与背景的对比度低等问题给现有的目标检测算法带来了挑战。为此,本文提出了一种基于自适应排序融合注意力网络(ASFANet)的端到端缺陷检测框架,其中设计的自适应双边分支子网络(ABBN),能够针对不同类型缺陷提取跨尺度、跨层次的视觉特征,并进行自适应地加权融合;其中设计的排序融合注意力子网络(SFAN)可在抑制背景的同时凸显目标区域,更好地提取微小微弱目标的视觉特征。实验结果表明,提出的方法在实际采集的带钢缺陷数据集上达到了83.2%mAP,较Faster-Rcnn、Libra-Rcnn和YOLOV3分别有8.5%、11.5%和17.8%的提升,尤其在划痕,白点缺陷上检测效果有较大提升。更多还原

【Abstract】 The complex and changeable forms of strip surface defects, vastly different scales, and low contrast between defects and background have brought challenges to machine learning algorithms. Therefore, we propose an end-to-end defect detection framework based on the Adaptive Sort Fusion Attention Network(ASFANet). Firstly, in order to solve the problem of feature imbalance in the general object detection model, we design an Adaptive Bilateral-Branch Network(ABBN), which can extract cross-scale and cross-channel visual features for different defects, and can adaptively weighted fusion to obtain balanced semantic features;In addition, in order to better extract the visual feature of tiny and weak defects, we design a Sort Fusion Attention Network(SFAN), which can highlight the features of the target area while suppressing the background. The experimental results show that the proposed method has reached 83.2% mAP on the actual collection of strip steel surface defect datasets. Compared with Faster-Rcnn, Libra-Rcnn and YOLOV3, the mAP has 8.5%, 11.5% and 17.8% improvement, especially on scratches and white spot.更多还原