基于YOLOX-αSMV的带钢材料表面缺陷检测算法

YOLOX-αSMV Algorithm for Surface Defect Detection of Strip Steel Material

【目的】针对YOLOX算法在钢材表面缺陷检测中特征提取不充分、多目标缺陷检测能力较弱等问题,提出改进损失函数的多维度特征融合带钢材料表面缺陷检测算法。【方法】首先,在Backbone部分应用SPP_SF保留多尺度特征信息,提高分类精度。其次,在Neck部分加入多维度特征融合模块MDFFM,将通道、空间、位置信息融入特征向量中,加强算法的特征提取能力。最后,引入Varifocal Loss和α-CIoU加权正负样本,提高预测框的回归精度。【结果】实验结果表明,YOLOX-αSMV在NEU-DET数据集中的mAP@0.5:0.95达到了47.54%,较YOLOX算法提高了3.43%。【结论】算法在保持检测速度基本不变的情况下,对模糊缺陷和小目标缺陷的识别、定位能力明显提升。

【Objective】In order to solve the problems of insufficient feature extraction and weak ability of multitarget defect detection of YOLOX algorithm in steel surface defect detection,a multi-dimensional feature fusion strip material surface defect detection algorithm based on improved loss function is proposed.【Method】First of all,apply SPP_SF to the Backbone part to retain multi-scale feature information and improve classification accuracy.Secondly,the multi-dimensional feature fusion module MDFFM is added in the Neck part to integrate the channel,space and position information into the feature vector to strengthen the feature ex-traction ability of the algorithm.Finally,the introduction of Varifocal Loss andα-CIoU is weighted with positive and negative samples to improve the regression accuracy of the prediction box.【Result】The experimental results show that YOLOXαSMV in NEU-DET data set mAP@0.5:0.95 reaches 47.54%,which is 3.43%higher than YOLOX algorithm.【Conclusion】The algorithm significantly improves the recognition and localization of fuzzy defects and small target defects while keeping the detection speed basically unchanged.