Domain knowledge driven deep unrolling for rain removal from single image

Rain is a common weather that seriously affects the performance of outdoor computer vision applications. The quality of images taken in such weather is very poor. There are several popular methods for the removal of rain streaks from images; one such method is based on physical models and mathematical optimization, and another method is based on deep learning. However, these methods have their own shortcomings. The optimization-based method is complex, but the result is general. In the deep-learning-based method, some details of the background images are lost through a deep network. In this study, we developed a ResNet and denoising algorithm embedded in the ADMM framework as the background/rain prior. ResNet was trained using synthetic rainy/clear background image pairs as the training data. Then, we divided the images taken in rainy weather into parts with a rainless background and those with the rain streaks. The experiments revealed that the PSNR value of the derain results obtained using a combination of a residual network and the ADMM algorithm was approximately 3% higher than that of the other rain-streak removal algorithms. Moreover, the detailed images obtained were considerably clearer than the details obtained from other rain-streak removal algorithms, and the image quality was better.