Loss Balance Aware Algorithms for Single Image Deraining Network on Mobile Devices

Despite many recent advances in single-image deraining, a deraining neural network affordable for resource-limited or mobile devices remains desirable. Unlike the methods relying on enormous network architectures and careful hyperparameter tuning, we introduce a lightweight model integrated with a streamlined deployment schematic and dynamic weighting algorithms to achieve efficient and automated single-image deraining on mobile devices. At first, we construct the network and associate it with multiple loss functions to efficiently and effectively capture expected background textures and high-frequency oscillations caused by rain streaks. To improve usability on handheld instruments, we simplify the deployment process by omitting some time-consuming offline operations. Importantly, we present two dynamic weighting algorithms to automatically synergize varied optimization goals. These algorithms prevent our framework from tedious manual regulation, which enhances the flexibility our model on portable devices. We conduct qualitative and quantitative evaluations on synthetic and real datasets of rainy images. The experimental results demonstrate that our model can be easily deployed on mobile devices and outperforms other state-of-the-art approaches in robust rain removal, even for real images captured in heavy rain.