HNN: Hierarchical Noise-Deinterlace Net Towards Image Denoising

Amogh Joshi, Nikhil Akalwadi, Chinmayee Mandi, Chaitra Desai, Ramesh Ashok Tabib, Ujwala Patil, Uma Mudenagudi; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 3007-3016

In this paper we propose a hierarchical framework for image denoising and term it Hierarchical Noise- Deinterlace Net (HNN). Image denoising techniques aim to recover clean images from noisy observations by reducing unwanted noise and artifacts to enhance the clarity and introduce spatial coherence. Images captured during challenging scenarios suffer from granular noise inducing fine- scale variations in the image which occur due to the limitations of imaging technology or environmental conditions. This granular noise can significantly degrade the quality of the image making it less useful for applications like object detection image restoration/enhancement face detection and image super-resolution. From literature we infer learning global-local features significantly contribute in reducing unwanted noise and artifacts within images. Typically researchers rely on residual learning Generative Adversarial Networks (GANs) and Attention Mechanisms to learn global-local features. However these methods face challenges such as vanishing gradients limited generalization of generators in GANs lack of global context aware- ness and computation complexity in attention mechanisms leading to drop in performance. Towards this we propose a hierarchical framework to process both global and local information across distinct levels of hierarchy. More specifically we propose a hierarchical encoder-decoder network with a distinct Global-Local Spatio-Contextual (GLSC) block for learning of fine-grained features and high-frequency details in an image. The proposed frame- work improves image denoising as it allows the model to capture and utilize information from different scales ensuring a comprehensive understanding of the image content. We demonstrate the efficacy of proposed HNN framework on benchmark datasets in comparison with state-of-the-art methods with 5% (? in dB) increase in performance.