基于YOLOX的金属表面缺陷检测

Metal surface defect detection based on YOLOX

为提高金属产品在加工制造过程中的检测精度与速率,提出一种基于深度学习的检测方法,运用经改进的YOLOX网络模型检测金属表面缺陷。融合丰富上下文的骨干网络模块C2f_COT、优化的空间金字塔池化New_SPP、混合域注意力机制CBAM和自适应特征融合算法ASFF,选用高效的EIOU损失函数。改进之后,对金属缺陷特征信息的提取能力得到提升,特别是针对形状复杂且大小不一的缺陷。在NEU-DET数据集上,模型精度mAP值达77.16%,比原始模型提高3.05%,速率FPS达到每秒73帧。精度和速率均可满足实际工作中的金属表面缺陷检测需求,显著提升检测效率。

To enhance the detection accuracy and speed of metal products during the manufacturing process,a deep learning-based detection method was proposed,utilizing the improved YOLOX network model to detect defects on metal surfaces.The context-rich backbone network module(C2f_COT),the optimized spatial pyramid pooling module(New_SPP),the hybrid domain attention mechanism CBAM,and the adaptive spatial feature fusion algorithm(ASFF)were integrated,while the efficient EIOU loss function was employed.Following these improvements,the ability of model to extract feature information from metal defects is elevated,especially for defects of complex shapes and varying sizes.Tested on the NEU-DET dataset,the precision of model,measured by mAP,reaches 77.16%,which shows a 3.05%improvement over that of the original model,and the speed reaches 73 frames per second(FPS).Both the precision and speed meet the requirements for real-time metal surface defect detection,significantly enhancing the detection efficiency.