基于改进Faster R-CNN的高铁扣件弹条缺陷检测

Fastener clips defect detection based on improved Faster R-CNN in high-speed railway

摘要:

针对复杂光照环境导致的高铁扣件弹条缺陷检测困难问题，提出一种基于改进Faster R-CNN的弹条缺陷检测方法。通过多层卷积神经网络提取缺陷特征图，提高网络对缺陷特征的关注程度，降低对复杂光照环境干扰的影响；设计区域候选网络生成候选区域，并根据候选区域进行池化，在特征图中提取相对应的具体缺陷位置；利用区域候选网络的全连接网络层计算获得缺陷的具体类别与精确位置，得到最终的检测结果。所提算法可充分抑制光照环境干扰影响，显著增强缺陷特征的表征能力；简化了图像预处理环节，降低了对原始图像成像质量的要求。实验结果表明，所提算法能够实现对高铁扣件弹条缺陷的有效检测。与现有算法相比，具有较高的精确度和较强的鲁棒性，运算效率也得到显著提升。

Abstract:

In response to the difficulty in detecting defects in high-speed rail clip springs caused by complex lighting environments, an improved Faster Region Convolutional Neural Networks(R-CNN)-based defect detection method for clip springs is proposed. By extracting defect feature maps through multi-layer convolutional neural networks, the network's attention to defect features is enhanced, and the impact of interference from complex lighting environments is reduced. A region proposal network is designed to generate candidate regions, and based on these regions, pooling is performed to extract the corresponding specific defect locations in the feature maps. The fully connected layers of the region proposal network are employed to calculate the specific categories and precise locations of defects, yielding the final detection results. The proposed algorithm can fully suppress the interference of lighting environments, significantly enhance the representation ability of defect features, simplify the image pre-processing stage, and reduce the requirements for the quality of the original image. Experimental results show that the proposed algorithm can effectively detect defects in high-speed rail clip springs, and compared to existing algorithms, it has a higher accuracy, stronger robustness, and significantly improved computational efficiency.