关键词: Transformer, 图像去噪, 全局特征, 局部特征, 特征解耦

Abstract: Abstract: Current Transformer-based algorithms focus on capturing the global features of images, while ignoring the crucial role of local features in restoring image detail information. To address this issue, an image denoising algorithm based on local and global feature decoupling is proposed, which includes two multi-scale branches based on hybrid Transformer block(HTB) and a single-scale branch based on convolutional neural network(CNN), aiming at combining HTB's powerful global modeling capabilities with CNN's local modeling advantages, and yielding outputs with enriched contextual information and precise spatial details. HTB employs self-attention mechanisms to adaptively model spatial and channel-dimensional dependencies to activate a wider range of input pixels for reconstruction. Given the potential information conflicts across different branches, a feature transfer block is designed to facilitate cross-branch propagation of global features and suppress low-frequency information, thereby ensuring collaborative interactions among the branches. Compared to the Transformer-based denoising algorithm (Uformer) on the real-world image dataset DND, the proposed algorithm achieves an improvement of 0.14dB in peak signal-to-noise ratio (PSNR). For the synthetic image dataset Urban100, the proposed algorithm achieves an average PSNR enhancement of 0.41dB compared to the multi-stage denoising algorithm (MSPNet). Experimental results demonstrate that the proposed algorithm effectively removes image noise and reconstructs finer texture details.

Key words: Keywords: Transformer, image denoising, global feature, local feature, feature decoupling