MIINet: a multi-branch information interaction network for few-shot segmentation

Mainstream Few-Shot Segmentation (FSS) methods leverage the prototype concept, utilizing masked average pooling to capture the high-dimensional centroid of a semantic class. However, the neglect of higher-order statistical information of the foreground targets and the ambiguous semantic of the background present a bottleneck in FSS. To tackle these challenges, we introduce a Multi-Branch Information Interaction Network (MIINet). It utilizes features from diverse origins and levels to comprehensively absorb foreground details while augmenting the divergence between foreground prototype and background prototype in feature space. In the query auxiliary branch, we develop the Structural Information Expansion module, which simulates the process of human learning in visual constancy. It employs mid-level features to calculate the structural difference between support futures and query futures, then it reconstructs the support image to extract additional foreground information based on the structural difference. In the support auxiliary branch, we present the Background Filtering and Activation module, which utilizes the high-level features from the support set to purify the background prototype, thereby enhancing its discriminative ability. Furthermore, we propose the Query Feature Enhancement module to activate the query features. We conduct a series of experiments on the PASCAL- \(5^{i}\) $5^i$ and COCO- \(20^{i}\) ${20}^i$ datasets. MIINet achieves satisfactory performance compared to the state-of-the-art approaches, validating the efficacy of our method.