基于改进YOLOX-S网络的金属表面缺陷图像检测
Image Detection of Metal Surface Defects Based on Improved YOLOX-S Network

本文为了实现金属零件制造和加工环节中表面缺陷的图像检测识别，提高金属零件加工流水线过程中不合格产品的图像检测准确率，提升金属缺陷检测过程中的设备自动化水平，解决人工检测中易疲劳、检测效率低、检测精度低、主观性强和无法适应大批量高质量零件的检测问题。提出一种基于改进YOLOX-S的卷积神经网络金属表面缺陷图像检测模型。该模型在原有的YOLOX-S的网络基础上引入注意力机制模块，优化模型参数，改进边界框位置回归损失函数与置信度损失函数，构建和实现对金属表面常见缺陷种类的图像检测模型。改进后的YOLOX-S模型能更快地收敛，添加边界框位置回归损失函数和注意力机制的模型，mAP值由94.23%提升为96.14%，准确率提升效果最好，同时仅增加较少的推理时间。与YOLOX-S模型以及添加不同改进方法模型作对比分析表明基于边界框位置回归损失函数和注意力机制改进的模型综合识别效果最好，能满足金属表面缺陷图像检测的要求，减少金属零件生产过程中的缺陷产品流出。
In order to realize the image detection and identification of surface defects in the manufacturing and processing of metal parts, improve the image detection accuracy of unqualified products in the process of metal parts processing assembly line, improve the automation level of equipment in the process of metal defect detection, and solve the problems of easy fatigue, low detection efficiency, low detection accuracy, strong subjectivity and inability to adapt to the detection of large quantities of high-quality parts in manual detection. A convolution neural network model based on improved YOLOX-S is proposed for metal surface defect image detection. The model changed the structure of the unit module and introduced the attention mechanism module based on the original YOLOX-S model, which optimized the model parameters and improved bounding box position regression loss function and confidence loss function, and then the image detection model of common defects on the metal surface was constructed and realized. The results showed that the loss of improved YOLOX-S model could converge faster, the model with bounding box position regression loss function and attention mechanism added, its mAP increased from 94.23% to 96.14%, and the accuracy improvement effect is the best, while only a small amount of reasoning time is increased. Compared with the YOLOX-S model and the model with different improved methods, it is shown that the improved model based on bounding box position regression loss function and attention mechanism has the best comprehensive recognition effect, which can meet the requirements of metal surface defect image detection and reduce the outflow of defective products in the production process of metal parts.