CoT-YOLOv8: Improved YOLOv8 for Aerial images Small Target Detection

Detecting small targets in aerial images using unmanned aerial vehicles is an important research direction in the field of object detection and a highly challenging task. However, existing object detection methods often suffer from high miss rates and false alarm rates in the task of detecting targets in aerial images. To address this issue, we propose an algorithm called CoT - YOLOv8 to improve small target detection in aerial images. Firstly, we add an additional detection layer to the YOLOv8 algorithm to enhance the detection capability for small target objects. Secondly, we insert multiple Convolutional Block Attention Module (CBAM) into the Backbone network to focus more on useful information, thereby improving the detection capability in complex scenes. Additionally, we replace the standard convolutional network in the Backbone network with a Dynamic Convolution Module (DCN), enabling the model to better adapt to geometric variations of the targets. Finally, we introduce the Contextual Transformer module into the Head network, allowing the model to utilize contextual information to assist in object detection and further improve the detection accuracy. The improved algorithm shows an increase of 7.7%, 7.2 %, and 8.7% in precision (P), recall rate (R), and average precision (IOU-O.5) respectively. This indicates that the CoT - YOLOv8 algorithm has better generalization capability and higher detection accuracy compared to the original YOLOv8 in aerial small target scenarios.