MDIINet: A Few-Shot Semantic Segmentation Network by Exploiting Multi-dimensional Information Interaction

The fundamental goal of Few-Shot Segmentation (FSS) is to achieve accurate segmentation of irrelevant categories by training the network on a very limited support set of relevant classes. This task requires deep mining and efficient utilization of informative training data, as well as the extraction of fine-grained correspondences between the support set and the query set. To address these challenges, we propose a Multi-Dimensional Information Interaction Network (MDIINet) that explores different dimensions of information correlation and different scales of information correlation. It extracts fine-grained correlation information at different levels from transformed convolutional layers in various dimensions, as well as coarse-grained correlation information at different levels from the concatenation process of transformed scales. This method employs a pyramid structure to differentiate information from coarse to fine in different dimensions, and performs a series of super-correlation operations to integrate fine-grained and coarse-grained information from each level to obtain the final feature output. The effectiveness of the proposed method has been validated on the PASCAL-5i and FSS-100 benchmark few-shot segmentation datasets, and the experimental results show significant performance improvements over state-of-the-art methods.