Shared multi-branch framework and enhanced feature fusion for few-shot semantic segmentation

Traditional semantic segmentation relies on massive annotated datasets, which are often prohibitively expensive in specialized fields such as medical or satellite imagery. This paper proposes an enhanced feature extraction framework for self-support few-shot semantic segmentation that overcomes the challenge of data scarcity. Unlike conventional single-backbone designs, the approach utilizes a multi-branch shared network that synergistically integrates ResNet, EfficientNet and CLIP, capturing a robust spectrum of local details and global semantic priors. A weighted mask feature extraction module dynamically focuses the model’s attention on precise foreground regions, reducing the impact of background noise in low-data regimes. An enhanced feature fusion strategy bridges high-level semantic features and low-level structural details while maintaining a compact parameter footprint. Experimental results demonstrate that the method significantly narrows the performance gap in data-constrained scenarios, achieving a mean Intersection over Union increase of 4.6% in one-shot and 4.5% in five-shot settings over the baseline, highlighting its effectiveness in extracting discriminative representations from minimal support information.