Consistency is Key to Image Denoising
Without noise, image processing and pattern recognition would be a fait accompli! Image denoising remains a fundamental challenge in practical image processing applications. Deep neural networks (DNNs) have set new performance benchmarks over classical techniques by using supervised learning to predict either the clean image or the noise exclusively. This paper brings both paradigms on the same footing and examines them from the minimum meansquared error (MMSE) estimation perspective, which requires that their outputs must satisfy a linear constraint, referred to as the consistency criterion. We show that jointly optimizing the image prediction and noise prediction models by enforcing the consistency criterion bridges the gap between the two paradigms and improves denoising performance across multiple settings. Experiments show up to 0.49 dB Peak Signal-to-Noise Ratio (PSNR) gain on CBSD68, Kodak24, McMaster, Urban100 datasets, and 0.46 dB on SIDD dataset, across diverse architectures like DnCNN, SwinIR, CycleISP, and Restormer, under both synthetic and real-world noise.