High-Speed Object Detection in Near Space Based on Improved YOLOv8s

Aiming at the problem of low accuracy and weak robustness of small target detection in high-speed near-air environment of traditional network models, a CDGH-YOLO algorithm model based on YOLOv8s optimization and improvement is proposed. The model reduces the number of parameters and the amount of calculation by integrating lightweight convolution modules (Ghost convolution and DW convolution); at the same time, the recursive gated convolution module (C3HG) is adopted to greatly improve the model's detection ability for small objects; in addition, by adding the input-output attention module (CBAM), the efficiency and robustness of the overall small target feature recognition of the model in special scenarios are improved; finally, by replacing the full intersection-over-union (CIoU) loss function with the WIoU v3 (Wise-IoU version 3) loss function, more attention is paid to the normal quality anchor box of small and medium-sized targets, thereby improving the detection efficiency of small samples. Experimental results show that the CDGH-YOLO algorithm model is superior to the baseline model YOLOv8s in terms of target recognition accuracy, recall rate, and mAP value. The versatility of the model was tested using the NWPU VHR-10 dataset. The results showed that the target recognition accuracy increased by 9.5 percentage points, and the recall rate and [email protected] increased by 2.6 and 2.7 percentage points respectively. These performance indicators are significantly better than the comparison algorithms, meet the requirements of edge computing devices, and have practical application value.