Abstract:In response to challenges in current image semantic segmentation tasks, such as inadequate small-object segmentation accuracy, blurred boundaries, and semantic confusion among similar categories, this study proposes a semantic segmentation method based on channel attention and progressive frequency-aware fusion. Building upon the SegFormer architecture with MiT-B0 as the backbone network, this approach dynamically adjusts channel weights to enhance the representation of key semantic features, while reducing sensitivity to noise during multi-scale feature fusion. At the decoder stage, an adaptive high-low pass filtering strategy and Carafe upsampling operator are employed to upsample high-level features smoothly and simultaneously enhance high-frequency details in low-level features. This leads to a progressive frequency-aware feature fusion strategy, which improves small-object segmentation accuracy, refines boundary clarity and strengthens semantic discrimination capabilities between similar categories. Experimental results demonstrate that with an increase of only 0.15M parameters, the proposed method achieves mIoU scores of 76.67% on Cityscapes and 37.85% on ADE20K datasets. It effectively enhances small object recognition, boundary restoration, and discrimination of easily confused categories, achieving a favorable balance between performance and lightweight efficiency.