Small Target Detection Algorithm Based on SDBNet and Multiple Detection Heads

Despite the great success of deep learning in object detection at present, its performance and efficiency for small target detection in UAV images are still unsatisfactory. In order to improve the performance of small target detection, which proposes an algorithm MS-YOLO based on the YOLOv8 algorithm to improve the backbone network and the detection head. Firstly, the SPD (Space-To-Depth Convolution), which is a feature extraction layer suitable for low-resolution images and small target images, is integrated into the shallow layer of the backbone network, which does not use stepwise convolution and pooling to better retain the small target feature information. This layer does not use step-width convolution and pooling, which can better retain the small target feature information. Secondly, a detection header specialized for the small target feature layer is added. The improved algorithm achieves state-of-the-art detection results on the VisDrone dataset with an average precision (AP) of 46.5 and 28.0, respectively, and a model parameter count of 16.8M, which improves the detection accuracy while maintaining lightweight.