Zero-shot industrial image denoising with lightweight network

Industrial images are often captured under full-time and full-weather conditions, leading to inevitable noise during the imaging process, which can impact subsequent detection algorithms. In recent years, image denoising with neural networks has been the rapid development. However, training such networks typically requires a large dataset, which is scarce in publicly available industrial image databases. In this paper, we propose a novel approach termed Zero-Shot Industrial Image Lightweight Denoising (ZSILD) network, which effectively denoises single noisy industrial image without the need for datasets. First, we sample the paired neighbour pixels of a random noisy industrial image, which are then utilized to train a lightweight denoising network. Second, we design a lightweight depthwise convolutions network based on bottleneck residual structure with shortcut connections. Finally, this network is trained on the sampled pairs using a novel loss function aimed at enhancing denoising performance. Our experiments conduct on real-world industrial ambient noise demonstrate that our ZSILD method outperforms existing denoising techniques, all while requiring comparatively minimal computational resources.