Efficient and Fast Self-Supervised Image Denoising Based on Cyclic Multi-Variate Function

Recent advancements in neural networks have driven notable progress in image denoising. Nevertheless, conventional supervised approaches rely heavily on large datasets of paired noisy and clean images, which restricts their practicality. Meanwhile, self-supervised denoising methods that train directly on single noisy images have been developed but face limitations such as inefficiency during training, information loss, and reliance on noise modeling. The Cyclic Multi-Variate Function for Self-Supervised Image Denoising (CVF-SID) method offers notable improvements to address these problems but still faces challenges with high resource consumption and long training cycles. In this paper, we present an improvement to the CVF-SID method—Efficient and Fast Cyclic Multi-Variate Function (EF-CVF) that significantly reduces GPU resource consumption and accelerates convergence, enabling more efficient training of image denoising models using only noisy images. The key improvements include adjusting the number of convolutional layers, Combining Efficient Channel Attention with Squeeze-and-Excitation, we designed the Squeeze-and-Excitation and Channel Attention (SECA) mechanism, which effectively integrates the strengths of both methods to enhance feature representation., and introducing a Partial Convolution (PConv) module. Our method, in comparison with the original CVF-SID method, effectively reduces GPU load, improves model training speed, and achieves enhanced denoising performance. We provide a theoretical explanation of the method and validate its effectiveness through extensive experiments on real-world datasets.