基于E-YOLOX的实时金属表面缺陷检测算法

A real-time metallic surface defect detection algorithm based on E-YOLOX

针对现有基于深度学习的金属表面缺陷检测方法存在泛化能力差、检测速度低等问题,提出一种新的检测算法E-YOLOX。该算法采用新的特征提取网络ECMNet,并使用深度卷积减少网络参数;以线性逆瓶颈残差网络提升特征提取能力,在正向传播过程中保留更多高维张量内的流形分布于低维子空间的关键特征;以扩张跨阶段局部网络结构多样化神经网络的梯度流路径,使深层神经网络更高效地学习和收敛。同时,提出一种新的数据增强方法边缘Cutout,在训练过程中自适应生成掩膜覆盖图像的随机区域,提升网络的检测和泛化能力。实验结果表明,E-YOLOX-l在铝型材表面缺陷数据集AL6-DET上检测精度达到了77.2%的mAP,在钢材表面缺陷数据集GC10上检测精度达到了36.8%的mAP,较基准模型YOLOX-l分别提高3.6%和1.7%,同时参数量减少55%,计算量减少49%,检测速度达到57 FPS,提高了21 FPS。与相关算法对比,该算法取得较高的检测精度,且在精度和速度之间达到较好的均衡。

For metallic surface defect detection,a novel algorithm E-YOLOX was proposed to address the shortcomings of current methods,such as poor generalization ability and low detection speed.The algorithm utilized a new feature extraction network,ECMNet,which employed depth convolutions to reduce the parameters and computational cost of the network.The linear inverse bottleneck residual network was in use to enhance the feature extraction capability,while preserving more key features that were manifold distributed in low-dimensional subspaces within high-dimensional tensors during forward propagation.Additionally,the extended cross-stage partial network structure diversified the gradient flow paths of neural networks,making deep neural networks learn and converge more efficiently.Moreover,a new data augmentation method edge Cutout was proposed,which generated adaptive masks covering random regions of the image during the training process,enhancing the detection and generalization ability of the network.The experimental results demonstrated that E-YOLOX-l achieved 77.2%mAP in detection accuracy on the aluminum profile surface defect dataset AL6-DET and 36.8%mAP on steel surface defect dataset GC10-DET,which was 3.6%and 1.7%higher than the baseline algorithm YOLOX-l.At the same time,the number of parameters was reduced by 55%and the computational cost was reduced by 49%.The detection speed was 57 FPS,an increase of 21 FPS.Compared with other related algorithms,the new algorithm achieved a higher detection accuracy and a better balance between accuracy and speed.