From CNNs to Adaptive Filter Design for Digital Image Denoising Using Reinforcement Q-Learning

Multi-modal image acquisition techniques have allowed digital images to penetrate domains from micro-scale medical imaging to mega-scale satellite imaging. For postprocessing, deep learning techniques have widely been used for image denoising and artifact suppression. However, far little work has been done to summarize their effectiveness concerning adaptive filter design, e.g., salt and pepper noise, stochastic Poisson, or additive white noise. Because different images, natural or urban, structured or unstructured scenes, and objects produce different types of noise, from the modality as well as from the imaging medium, devising a single method for all noise types is impractical. This paper proposes to use reinforcement learning (Q-learning) to adaptively design filters of a convolutional neural network (CNN). In contrast to the popular state of the art methods that use filter designs based on the noise model, CNN filters lack the power to do so. We have attempted to address this limitation of CNN by introducing a new modality of reinforcement learning for adaptive filter design. The qualitative and quantitative analysis of the proposed method is done and its efficacy is demonstrated using the following evaluation metrics: Peak Signal to Noise Ratio (PSNR), Contrast to Noise Ratio (CNR), and Structure Similarity Index (SSIM).