Efficient pavement bad location detection and repair is essential to prolong the use time of roads. However, traditional manual detection methods are extremely inefficient and can no longer meet the requirements of inspecting a large number of roads. When using deep learning technology for road disease detection, it is found that low-illuminance images will affect the detection accuracy due to low contrast. Therefore, before training and testing the deep learning model, the original image needs to be preprocessed to improve the image quality. First, bilateral filtering is used instead of Gaussian filtering to estimate the illuminance of the original image; Then the reflection component is get according to the principle of Retinex algorithm, and the reflection image is quantized; Finally, the image is subjected to illumination compensation. The results of comparative experiments display that the ours algorithm can retain the characteristic details of road diseases and eliminate the unevenness of the image brightness distribution while improving the contrast of the road image. © 2022 IEEE.