SODCNN: A Convolutional Neural Network Model for Small Object Detection in Drone-Captured Images

Drone images contain a large number of small, dense targets. And they are vital for agriculture, security, monitoring, and more. However, detecting small objects remains an unsolved challenge, as they occupy a small proportion of the image and have less distinct features. Conventional object detection algorithms fail to produce satisfactory results for small objects. To address this issue, an improved algorithm for small object detection is proposed by modifying the YOLOv7 network structure. Firstly, redundant detection head for large objects is removed, and the feature extraction for small object detection advances. Secondly, the number of anchor boxes is increased to improve the recall rate for small objects. And, considering the limitations of the CIoU loss function in optimization, the EIoU loss function is employed as the bounding box loss function, to achieve more stable and effective regression. Lastly, an attention-based feature fusion module is introduced to replace the Concat module in FPN. This module considers both global and local information, effectively addressing the challenges in multiscale and small object fusion. Experimental results on the VisDrone2019 dataset demonstrate that the proposed algorithm achieves an mAP50 of 54.03% and an mAP50:90 of 32.06%, outperforming the latest similar research papers and significantly enhancing the model's capability for small object detection in dense scenes.