基于Retinex模型和GTV的铁路货车铸件DR图像增强

DR image enhancement of railway freight car castings based on Retinex model and GTV

铁路运输已成为我国最重要的运输方式之一,铁路货车部件(摇枕、侧架等)的铸造缺陷(气孔、裂纹等)可能造成交通事故。由于铁路铸件自身的不均匀性会导致原始DR图像灰度不均,不易捕捉缺陷细节,因此需要对原始DR图像进行图像增强等操作检测缺陷。基于Retinex模型,同时考虑目标的光照强度和材质信息,将图像分解为光照图和反射图。基于高斯全变分(GTV)的保护图像边缘滤波器可以平滑光照图像,并且可以移除光照图的纹理细节,因此缺陷等细节就暴露在反射图中。对于铁路货车铸件的缺陷识别问题,结合Retinex和GTV方法,提出一种改进的Retinex增强模型,利用GTV和基于纹理感知的加权项对Retinex分解过程中的光照图和反射图进行正则化。此外,用交替优化算法求解该模型。最后,将能够反映图像缺陷细节的反射图作为最终增强图像。此方法得到的图像不仅保留了图像的结构信息,还明显暴露了图像的缺陷细节。实验结果表明,与其他已有的模型相比,该模型增强缺陷效果显著,此外,该模型提高了DR图像的信息熵、平均梯度等量化指标。与原图相比,增强后图像的信息熵提高了8%以上,平均梯度至少提高为原来的6倍。此方法提高了DR图像检测缺陷的能力,可以应用于铁路货车铸件等无损检测领域。

Railway transportation has become one of the most important transportation modes in China. Casting defects(pores, cracks, etc.) of railway freight car parts(such as pillow or side frame) may cause traffic accidents.Since the inhomogeneity of the railway castings could cause uneven gray scale of the original DR images, it is difficult to capture the defect details. Therefore, image enhancement and other operations should be performed on the original DR images to detect the defects. The model based on Retinex takes into account both the illumination intensity and material information of the target and decomposes the images into illumination and reflection ones.The protective image edge filter based on Gaussian Total Variation(GTV) can smooth the illumination images and remove the texture details of the illumination images so that the details of defects are exposed in the reflection image. For defect identification of railway truck castings, an improved Retinex enhanced model was proposed, which used GTV and texture perception based weights to regularize the illumination and reflection maps in the process of Retinex decomposition. In addition, an alternative optimization algorithm was used to solve the model. Finally, the reflection image reflecting the details of the image defect was used as the final enhanced image. The image obtained by this method not only retains the structure information of the image, but also obviously exposes the defect details of the image. The experimental results show that compared with other existing models, this model enhances the defects significantly. In addition, the information entropy and mean gradient of DR Images were improved. Compared with the original images, the information entropy of the enhanced image was increased by more than 8%, and the average gradient was increased by at least 6 times. This method improved the ability of DR image detection and can be applied to non-destructive testing of railway truck castings.