A Few-Shot Semantic Segmentation Network with Mixed Prototypes and Base Learner

Few-shot semantic segmentation methods aim to segment new object categories with limited labeled samples. Existing approaches struggle to extract comprehensive semantic matching information from target regions and mitigate noise from complex backgrounds. To address background noise, this paper proposes a novel few-shot semantic segmentation approach using prototype generation and base class learning. A pyramid pooling convolutional layer extracts rich prototype vectors by discerning differences between the core mask and label mask. Additionally, a base class learning module captures semantic information from base class objects during training, effectively suppressing interference from background regions and aiding in background semantic matching. Comparative and ablation experiments demonstrate the superior performance of the proposed model. Evaluation on benchmark datasets shows its effectiveness, with results on Pascal-5i dataset surpassing the baseline model, achieving an average intersection over union of 67.15% in one-shot scenarios.