NATCA YOLO-Based Small Object Detection for Aerial Images

Abstract

The object detection model in UAV aerial image scenes faces challenges such as significant scale changes of certain objects and the presence of complex backgrounds. This paper aims to address the detection of small objects in aerial images using NATCA (neighborhood attention Transformer coordinate attention) YOLO. Specifically, the feature extraction network incorporates a neighborhood attention transformer (NAT) into the last layer to capture global context information and extract diverse features. Additionally, the feature fusion network (Neck) incorporates a coordinate attention (CA) module to capture channel information and longer-range positional information. Furthermore, the activation function in the original convolutional block is replaced with Meta-ACON. The NAT serves as the prediction layer in the new network, which is evaluated using the VisDrone2019-DET object detection dataset as a benchmark, and tested on the VisDrone2019-DET-test-dev dataset. To assess the performance of the NATCA YOLO model in detecting small objects in aerial images, other detection networks, such as Faster R-CNN, RetinaNet, and SSD, are employed for comparison on the test set. The results demonstrate that the NATCA YOLO detection achieves an average accuracy of 42%, which is a 2.9% improvement compared to the state-of-the-art detection network TPH-YOLOv5.