Multi-stage cross-transformer-based CNN for image denoising

Deep convolutional neural networks have shown exceptional performance in eliminating spatially invariant noise, such as synthetic noise, from images. However, their effectiveness diminishes with real-world images, often requiring complex, multi-stage architectures. To overcome this challenge and improve the practical applicability of denoising algorithms, this paper introduces a novel model called cross-Transformer denoising convolutional neural network (CTCNet). CTCNet is built upon three core components: consecutive blocks (CB), parallel blocks (PB), and a filtering block (FB). The CB employs a ConvNeXt block to extract essential structural information, ensuring effective noise removal. To minimize information loss, PB integrates three heterogeneous networks to support multi-level feature interaction, enhancing adaptability to complex real-world scenes. Transformer blocks are strategically embedded within both CB and PB to further boost performance. A fusion Module aggregates outputs from the sub-blocks, promoting inter-channel feature interaction. Additionally, this study introduces Dual multilayer perceptron (Dual MLP) and Cross-Dual multilayer perceptron (Cross-Dual MLP) architectures, which extract complementary features, enhancing the model’s ability to capture pixel-level relationships and remove noise effectively. The final FB refines the processed image to generate a clean result. Experimental results demonstrate that CTCNet consistently outperforms several existing denoising techniques, achieving superior performance on both synthetic and real-world noise removal tasks.