Rethinking the structural similarity in small object detection

In the field of Small Object Detection (SOD), accurate classification and localization are crucial for detection performance. However, previous works have largely focused on region accuracy, with less attention given to spatial information. It makes the performance of the detector in small objects constrained. Therefore, we propose a novel object detection algorithm named space-aware small object detector (SASOD). Specifically, to enrich spatial features during the feature extraction process, we designed a Spatial-aware Convolution module (SAConv), which effectively captures and reconstructs the spatial structural information of images. Concurrently, we enhanced the model’s boundary perception capability by introducing a structural similarity penalty term in the loss function. In the experimental section, we employed the widely-used VisDrone2019 dataset to evaluate the performance of the SASOD. YOLOv10 is selected as the benchmark experiment, and the designed algorithm is fused to YOLOv10, compared with the baseline model, the fused model’s experimental metrics mAP on the VisDrone2019 dataset is improved by 2.46%. Through comparative experiments with existing state-of-the-art object detection algorithms in other datasets, we demonstrated SASOD’s significant improvement in boundary localization accuracy and overall detection performance.