MPDenoiseNet: Resource-Efficient Deep Learning Approach for Image Denoising

Image denoising is a crucial preprocessing step in various computer vision applications, yet remains challenging due to the diverse nature of noise and the trade-off between denoising efficacy and computational cost. This paper introduces MPDenoiseNet, a multi-path, resource-efficient deep learning network for image denoising, designed to handle a range of noise types including Bernoulli, Poisson, Salt-and-Pepper, and Gaussian noise, while being more computationally efficient than the transformer based, noise type agnostic restoration models. MPDenoiseNet architecture integrates parallel paths of Selective Convolutional (SeConv) blocks, adept at denoising Salt-and-Pepper corrupted images, and novel Anisotropic Diffusion blocks, designed to smooth images while preserving structural details. The features extracted from these paths are fused and further refined by Transformer blocks, inspired by Restormer, to capture long-range pixel interactions. Finally, an asymmetric U-Net AutoEncoder with skip connections module serves as a post-processing stage to enhance perceptual quality. MPDenoiseNet manages to reduce dependence on computationally expensive transformer blocks, using significantly fewer of them than the current state-of-the-art models and replacing them with more specialized lightweight blocks. We demonstrate MPDenoiseNet denoising performance is competitive with state-of-the-art methods like Restormer, while exhibiting efficient resource utilization in terms of both runtime and memory consumption.