PID Controller-Inspired Model Design for Single Image De-Raining

Man Zhou, Fan Wang, Xian Wei, Rujing Wang, Xue Wang

Abstract: Deep learning based methods have achieved remarkable breakthroughs on the single image de-raining task. However, most of the current models are constructed by empirically designing black-box network architectures. These network architectures always lack sufficient interpretability, which limits their further improvements in de-raining performance. In this brief, inspired by the classical Proportional Integral Derivative (PID) controller and feedback mechanism in the automatic control community, we propose a novel de-raining network architecture to address the above issues. Specifically, since the PID controller can speed up the system convergence and eliminate the system steady-state error, this motivates us to mimic the signal processing flow of PID controller to provide an interpretable and reliable guideline on de-raining network designs. By casting the signal flow process in the PID controller as blueprints, we extend the PID controller to an almost parameter-free network module, named PID-IM with every component in the module one-to-one corresponding to each operation involved in PID controller. Equipped with PID-IM, our proposed network could efficiently explore and exploit the features of rain streaks in a recursive fashion. Extensive experiments on several benchmarks demonstrate that our method has best performance over other de-raining methods. In addition, by directly embedding our PID-IM into existing baseline networks, the de-raining performance can be significantly improved.