Authors: Zohaib Ali
Abstract: Image denoising is a foundational stage of the image signal processing (ISP) pipeline: its output quality bounds every downstream task, including demosaicing, compression, retrieval, and recognition. State-of-the-art deep denoisers (DnCNN, FFDNet, CBDNet) achieve strong quality but are typically designed and evaluated for offline, GPU-server settings, leaving embedded and real-time deployment as a secondary concern addressed only through post-hoc compression. This paper designs a denoiser to be lightweight from the outset rather than compressed after the fact. We propose a 3-layer residual convolutional network (5.9K parameters) with an auxiliary noise-level input channel (FFDNet-style conditioning), trained across a range of noise levels rather than a single fixed level. In a controlled study, we first show that a comparable single-noise-trained variant generalizes poorly outside its training noise level (PSNR drops from 28.6 dB at sigma=25 to below classical-filter performance at sigma=50). We then show that noise-level conditioning directly closes this gap: the conditioned model matches or exceeds Gaussian blur and median filtering across sigma in {10, 25, 50} without retraining, using three orders of magnitude fewer parameters than DnCNN. A channel-width ablation (C=16, 24, 32) further shows that quality does not increase monotonically with capacity under a fixed training budget, underscoring that training schedule — not just architecture size — is central to the lightweight-denoising design space. All results are produced by directly executing the accompanying code (no benchmark numbers are copied from other papers); we report exact scope, data, and hardware limitations in Section 6 alongside a concrete roadmap to full-scale benchmark evaluation (BSD68, Set12, SIDD).
