HASO-DETR: hybrid attention small object detection based on RT-DETR

Abstract

In unconstrained environments such as aerial and surveillance imagery, achieving precise detection of small objects continues to pose significant challenges in computer vision because of constrained pixel data and cluttered backgrounds, and high sensitivity to bounding box localization errors. Although existing detectors have made progress, they often suffer from the gradual loss of fine-grained features in deep networks and inefficient multiscale feature fusion. Motivated by these challenges, this study designs the HASO-DETR small object detection algorithm.The proposed framework features three key contributions: 1 The Hybrid Dilated Residual Attention Block (HDRAB) embedded in the backbone network, which employs multi-scale dilated convolutions and channel attention mechanisms to enhance and preserve critical features of small objects during downsampling stages; 2 A redesigned cross-scale feature fusion module, CCFM-S2, which utilizes the SPD-Conv operator for informationpreserving downsampling and explicitly integrates high-resolution shallow features (S2 layer), thereby infusing indispensable spatial details into the feature hierarchy for small targets; 3 The adoption of Employing the NormalizedWasserstein Distance (NWD) as a regression loss, we model bounding boxes as Gaussian distributions to effectively reduce sensitivity to minor deviations in bounding box coordinates. Extensive experiments on challenging benchmark datasets VisDrone2019 , demonstrate that HASO-DETR sets a new state-of-the-art performance, outperforming numerous mainstream detectors.