Single Image Rain Removal Using Convolutional Neural Network

Abstract

The visibility of the image is highly affected by the rain streaks. This can influence the performance of numerous visual tasks, for example, image enhancement, object tracking, recognition, surveillance and autonomous navigation. Process of recognition and removal of rain streaks is a quite complex and difficult task since there is no spatial-temporal content of rain streaks in a single image as compared to the dynamic video. This paper proposes an improved convolutional neural network (CNN) architecture to recognize and remove the rain streaks. Linear additive composite model is used for making rainy image model. Network is trained on the pre-processed image, which helps to enhance the learning of the network weights and training without huge increase in training data or computational resources. The experimental work shows that the CNN architecture successfully performed on both synthesized and real-world rainy images.