ADMNet: a small-sample semantic segmentation framework for the fusion of adaptive prototyping and diffusion modeling

The main challenge in the field of small-sample semantic segmentation is how to accurately realize the segmentation of target objects in query images with only a small number of labeled samples. Many existing approaches to this problem often require a compromise between preserving complex local semantic features and controlling computational complexity. To address these issues, a novel approach based on the Transformer architecture is proposed, which incorporates the advantages of the diffusion model. Firstly, an adaptive prototype learning mechanism is introduced, which enables the model to capture and represent feature differences between categories more accurately by dynamically adjusting the prototype vectors. Secondly, a Transformer decoder and a context mask generation module are added to improve the segmentation granularity. Finally, an efficient diffusion model with adaptive computation is utilized to simulate the data generation and denoising process to improve the model's ability to capture global features. Extensive experiments on PASCAL-5i and COCO-20i show that the model achieves significant performance improvement, proving its advantages in dealing with new classes of image segmentation problems.