Lightweight Reversible Network for Infrared Image Denoising

With the rapid advancement of deep learning, designing deeper and more complex neural networks has become a key approach to improving the accuracy of infrared image denoising. However, such network architectures are often accompanied by a large number of model parameters, making efficient deployment in resource-constrained environments, such as development boards, challenging. This study proposes a reversible neural network framework for infrared image denoising. The method involves separating the noise component from the infrared noise image during the forward propagation process, followed by reconstructing the denoised image in the backpropagation process. We further optimized the lightweight version of the network, successfully deploying it on a Ti60 FPGA and validating its superior performance in real-world scenarios.