Improved YOLOv5 Algorithm for Small Object Detection in Drone Images

The object detection in the context of drone is a hot topic in the field of computer vision in recent years. In response to the challenge of limited image feature information and the presence of numerous small and densely packed objects in drone-captured images, this paper proposes a novel feature fusion detection model, HTH-YOLOv5, based on YOLOv5. Firstly, we enhance the detection capability of small objects by adding a detection channel from high-resolution feature maps and propose a Hybrid Transformer Head (HTH) that incorporates a hybrid Transformer module, aiming to improve the network’s focus on small objects by fusing global and local feature information. Secondly, we introduce a Convolutional Attention Feature Fusion module (CA-FF) based on CBAM. This module dynamically adjusts attention weights for the allocation of original feature maps in both channel and spatial dimensions, aiming to enhance the feature extraction capability for small objects. Finally, to better capture global and contextual information, we introduce the Hybrid Transformer module into the backbone and enhance its original feature fusion method using the CA-FF module. Experiments on the Vis-Drone 2021 dataset demonstrate that, compared to the baseline YOLOv5s model, the improved model shows an increase of 7.2% in \(mAP_{50}\) and 6.3% in \(mAP_{75}\) . The model trained with an input resolution of \(1540\times 1540\) achieves an \(mAP_{50}\) of 57.1%, marking a 12.4% improvement over YOLOv5. The improved HTH-YOLOv5 achieves increased accuracy while maintaining a detection speed of 45 FPS, making it more suitable for small object detection in drone scenarios.