Feature Fusion And Coordinate Attention For Small Target Detection

Abstract

CenterNet is an anchor-free and one-stage object detection network based on the keypoint. It has the advantages of simple network structure, fast detection speed and high accuracy. Due to the scarcity of small targets characteristics, insufficient low-level semantics information and lack of high-level position information, CenterNet faces great challenges in the process of small target detection and the performance is not ideal. Aiming at the above issues, we use ResNet as the feature extraction network and insert coordinate attention mechanisms into the residual block to make full use of the captured position information. At the same time, the Feature Pyramid Network is introduced, and asymmetric contextual modulation is used for feature fusion between high-level and low-level features, so as to encode semantic information and spatial details more abundantly. Through experimental comparison and analysis, the improved model improves the detection accuracy of small targets by 9.8% on the dataset compared with the original CenterNet, which proves the effectiveness and robustness of the improved model.