GGBDCA:Scene graph generation based on Global Gradient Balanced Distribution and Compound Attention

Scene Graph Generation (SGG) is a key task in computer vision, aimed at automatically extracting objects and their relationships from images. Despite significant advances in SGG, the challenge of long-tail distribution continues to impede accurate prediction of rare relationships. To address this, we propose a novel SGG method based on Global Gradient Balanced Distribution and Compound Attention (GGBDCA). First, we introduce a Transformer-based framework that leverages a compound attention mechanism to extract detailed features. To tackle the long-tail identification problem, we propose a Global Gradient Balanced Distribution (GGBD) algorithm, which converts the long-tail issue into a multi-objective optimization problem. Through Gradient Balance Grouping (GBG), similar categories are clustered to enhance attention on rare classes. Then, the Multi-Gradient Descent Algorithm (MGDA) is employed to solve the multi-objective optimization, while the Adaptive Calibration Function (ACF) dynamically adjusts classification scores to improve the model’s generalization. These three core modules—GBG, MGDA, and ACF—work together to balance learning between head and tail classes, focusing more on rare categories and boosting overall performance.