<p>Image denoising plays a crucial role in computer vision by enhancing image quality and offering more reliable data for subsequent analysis. However, existing deep learning-based denoising methods that stack multi-layer convolutional networks often struggle to effectively integrate multi-scale information, leading to suboptimal performance when confronted with complex noise patterns. Additionally, conventional architectures have difficulty capturing irregular feature information and effectively modeling long-range dependencies within images. This paper proposes a channel hourglass guided dual-path network for image denoising (CHDNet). Its core backbone consists of two key modules: the channel attention enhanced hourglass feature distillation block and the dual-path feature enhancement block. The former incorporates a channel attention mechanism into the hourglass structure after the upsampling layer. This enables the network to recalibrate the importance of different feature maps and better extract rich multi-scale information. Meanwhile, the latter leverages a graph Transformer to capture long-range dependencies and uses deformable convolutions to handle irregular local features, enhancing the model’s representation ability and denoising performance. Experimental results reveal that CHDNet performs well in both quantitative and qualitative evaluations. Meanwhile, it maintains a moderate level of algorithmic complexity, highlighting the method’s practicality and efficiency.</p>