Generative Adversarial Network For Rain Removal And Image De-Raining System Implementation

As market growing and hardware acceleration for deep learning,computer vision algorithms have achieved outstanding performance and got widespread use in real world scenario,while they can be degraded by raining weather in outdoor scenes.Rain images restoring is a solution to avoid this problem.In this paper we propose an image de-raining approach based on generative adversarial network.Existing methods for single image rain removal are always not powerful enough or take up much time in calculation.Referring the requirement of performance and time consumption,we propose our own network structure.Perceptual loss function is introduced in train phase so that sematic information can also be restored in image processing,and this information is crucial in any computer vision application.We also explore weakly supervised learning as artificial dataset cannot simulate rain scenes in a realistic way.Our image de-raining system contains two main modules:offline training and online application.Offline module is responsible for network training and adjusting network model according to online vision task.Online module works on web transmission and image processing.This module can also make a flexible change for different computer vision tasks.We take object detection task in rain environment as a case to test and verify each module of our system.Our method achieves better results in objective indexes compared to other rain removal methods.The system also meets the need of real-time image processing and tasks far less computation resources than other methods.