U-Net-Based Image Segmentation Network with Multi-Scale Feature Fusion and Attention Enhancement

Abstract

Image segmentation, as a critical pixel-level prediction task in computer vision, demands higher spatial consistency and boundary accuracy in applications such as medical imaging, remote sensing interpretation, and industrial inspection. Addressing issues like insufficient scale adaptability and uneven feature response distribution in encoder-decoder deep learning segmentation models, this study constructs an improved image segmentation network based on the U-Net backbone. This approach preserves end-to-end training and spatial alignment while introducing a multi-scale feature fusion module to enhance semantic expression across varying receptive fields. An attention mechanism embedded in the decoding stage enables adaptive weighting of features across both channel and spatial dimensions. Additionally, a composite loss function is employed to jointly constrain pixel discrimination and region overlap relationships. Performance evaluation and ablation studies were conducted by comparing multiple typical segmentation models under unified experimental settings. Experimental results demonstrate consistent improvements across metrics including IoU, Dice, and Recall. Ablation studies validate the synergistic contributions of multi-scale fusion and attention mechanisms. Visualization further reveals enhanced stability in maintaining object boundary continuity and preserving fine-grained structures. These findings indicate the model achieves more reliable pixel-level segmentation in complex scenarios, offering a viable deep learning solution for high-precision image understanding tasks.