【摘要】 针对铁路C4检测车夜间采集图像由于补光不足造成的对比度偏低、亮度不足、明暗不均，严重影响后续故障识别及人工校验问题，提出一种基于深度强化学习的局部零部件区域像素级接触网图像质量增强方法。为克服接触网检测图像无效背景区域像素占比大、传统全局图像增强算法受背景干扰效果差的问题，基于PixelRL强化学习理论，设计针对接触网图像统计特征的多Agent动作组合；为得到配对训练数据集，利用EnlightenGAN算法，在对抗中生成增强的接触网配对图像；以局部像素点为操作对象，实现迭代过程中像素操作的可视化，整体算法具有较高的可解释性。实验结果表明：本方法可以在实现接触网图像增强的同时避免检测图像背景区域干扰，克服局部零部件区域像素特征信息丢失，不同区域图像亮度均衡性优于传统算法。更多还原

【Abstract】 To address low contrast, insufficient brightness and uneven brightness caused by insufficient light supplement in the catenary images collected by the C4 railway inspection vehicle at night, which seriously affects the subsequent fault identification and manual verification, a pixel-level catenary image quality enhancement method was proposed based on deep reinforcement learning for the local components area. To overcome the problem of large pixel share of the invalid background area of catenary detection images and poor effect of traditional global image enhancement algorithm due to background interference, based on PixelRL deep reinforcement learning theory, a multi-agent action combination was designed for the statistical features of catenary images. In order to obtain a paired training data set, the EnlightenGAN algorithm was used to generate enhanced paired catenary images in the confrontation. By using local pixels as the operating object, the visualization of pixel operations in the iterative process was achieved, proving the high interpretability of the overall algorithm. The experimental results show that this method can achieve the catenary image enhancement while avoiding the interference in the background area of the detection image and overcoming the loss of pixel feature information in the local parts area. Moreover, the brightness balance of different regions is better than traditional algorithms.更多还原