CA-YOLO: Model Optimization for Remote Sensing Image Object Detection

Abstract

The CA-YOLO (Coordinate Attention-YOLO) model has been optimized for object detection in complex remote sensing images, addressing key issues faced by algorithms that detect multiple objects. These issues include weak multi-scale feature learning capabilities and the challenging trade-off between detection accuracy and model parameter complexity. The CA-YOLO model, built on the framework of YOLOv5, incorporates a lightweight coordinate attention module in the shallow layer to improve detailed feature extraction and suppress redundant information interference. Additionally, a spatial pyramid pooling-fast with a tandem construction module is implemented in the deeper layer. The model employs a stochastic pooling strategy to fuse multi-scale key feature information from low-level to high-level layers, reducing the number of model parameters while improving inference speed. We optimized the anchor box mechanism and modified loss function to improve the ability of the model to detect objects of different sizes and scales. Results show that the CA-YOLO model outperforms the original YOLO in terms of multi-object detection accuracy, with an average [email protected] accuracy improvement of 4.8% and [email protected]:0.95 accuracy improvement of 3.8%. Additionally, the CA-YOLO model demonstrates exceptional inference speed, averaging 125 fps, which reinforces its superiority in detection accuracy, generalization ability, and overall efficiency. Notably, these improvements were achieved while maintaining the same number of parameters and complexity as other models, making the CA-YOLO model an exceptional choice for various applications.

Keywords

• Object detection
• attention mechanism
• coordinate attention
• SPPF
• SIoU loss