Effective Small Object Detection in Remote Sensing Images Based on Improved YOLOv8 Network

Abstract

Small object detection has long been a challenging and prominent research area in computer vision. Driven by deep learning, small object detection has made major breakthroughs and has been successfully applied in fields such as national defense security, intelligent transportation, and industrial automation. In our research, we conduct a comprehensive analysis and improvement of the YOLOv8-n algorithm for object detection, and propose two new algorithms. The first algorithm adds an SE Attention module and a detection head for small objects in YOLOv8. Improvements include the following points: a) adding a detection head for small objects to enhance detection capabilities for small objects, b) adding an SE Attention module to the network to improve the detection capability of the model. The resulting algorithm improves the performance of YOLOv8 in small object detection on the remote sensing image DOTA-v2.0 dataset and reduces the interference of noisy data to a certain extent. The second model incorporates the CA attention mechanism into the YOLOv8n model and uses the SEResNeXtBottleneck detection header instead of the YOLOv8n header. Through experimental validation on the DOTAv1 dataset, our improved model demonstrates a more accurate ability to detect small objects, and the experimental results show that by adding the CA attention mechanism, the model is able to focus more effectively on key regions of the image, thereby improving detection accuracy.