Dual-Branch Dynamic Graph Convolutional Network for Robust Multi-Label Image Classification

Abstract

For the intricate task of multi-label image classification, this paper introduces an innovative approach: an attention-guided dual-branch dynamic graph convolutional network. This methodology is designed to address the difficulties faced by current models when handling multiple labels within images. By integrating multi-scale features, it enhances the retention of original category information and boosts the robustness of feature learning. Utilizing a semantic attention module, the study dynamically reweights feature categories in the training dataset, enhancing the network's capability to identify smaller objects and generate context-sensitive category representations. The effectiveness of the proposed model was evaluated using the MS-COCO2014 imagery dataset, demonstrating superior performance in critical metrics such as classification precision (CP), recall (CR), and F1 score (CF1), outperforming other state-of-the-art models. Furthermore, a cascaded classification structure was implemented to leverage the prior information from static images to inform the processing of dynamic ones, and to utilize original image category data to augment label correlations, thereby enhancing overall classification accuracy.