Composite degradation image restoration

Abstract

In real-world road scenarios, adverse weather conditions and road pollutants present complex image degradation, a result of the interaction between multiple types of degradation also known as composite degradation. The majority of current SOTA methods do not serve as appropriate solutions as they only restore isolated degradation types. At the moment, OneRestore is one of the few models available that can handle composite degradation. However, it still cannot handle motion blur, a commonly occurring degradation found in traffic images due to both camera and object movement. In this project, we implement a modified version of OneRestore to handle more types of degradation suitable for road use. We also propose CDD-13 dataset, which is an extended version of the original CDD-11 dataset. The model loss function is modified to include Charbonnier loss, which further improved its current restoration ability. The modified OneRestore model exhibits exceptional performance when tested against the original CDD-11 dataset (29.10dB PSNR, 0.8850 SSIM), real-world benchmark datasets such as Snow100k-R (3.00 NIQE, 23.82 BRISQUE) and RTTS (4.56 NIQE, 31.74 BRISQUE), as well as other real traffic scenes.