Gated RedNet for image denoising

Deep learning has garnered widespread attention in the computer vision field, particularly for image denoising tasks. While Transformer-based models currently represent the state-of-the-art in denoising performance, this work investigates whether the architecture based on convolutional neural networks (CNN) can achieve comparable results. We propose GatedRedNet, a novel three-layer encoder-decoder network designed for image denoising. To enhance network depth while maintaining efficiency, GatedRedNet incorporates attention mechanisms within its residual blocks. Specifically, we incorporate the Efficient Channel Attention (ECA) module within all residual blocks, enabling effective recovery of edge and texture details with minimal computational overhead. To further improve global feature learning, we introduce a learnable gating factor at the encoder-decoder fusion stage, dynamically adjusting the weight ratio between their feature maps. These designs enable GatedRedNet to better capture local texture features while maintaining robust denoising capabilities. We evaluate GatedRedNet extensively on multiple benchmark datasets, demonstrating denoising performance comparable to state-of-the-art Transformer-based models and even surpassing them in certain aspects. Our results suggest that well-designed CNN architectures remain competitive in image denoising, offering an efficient alternative to Transformer.