Target Detection Algorithm for Aerial Images Based on Improved YOLOv8n

Abstract. In aerial images taken from the perspective of UAV, many ground objects are shown as small targets in the image because of the shooting height, and the small targets are small in size and not obvious in features, this makes the target detection algorithm is not well them checked out. Therefore, a new algorithm model based on YOLOv8n is designed in this paper, which introduces P2 detection header to detect objects with small spatial scale better. Cross-scale feature fusion (CCFM) neck network structure is introduced to make the feature scale more consistent after layer fusion. By integrating the attention mechanism of CPCA (Channel Prior Convolutional Attention) at the conclusion of the backbone architecture, we seek to improve the feature representation in both the channel and spatial aspects. The introduction of RFCBAMConv convolution allows for non-shared convolution kernel parameters during the convolution procedure, enabling more effective extraction of image features. In the end, Shape-IOU Loss was employed as the loss function to improve detection capabilities in multi-scale target detection situations by considering the influence of label frame dimensions on anchor frame regression. Experimental results demonstrate enhanced model in VisDrone2019 dataset achieved increases of 8.2 and 5.4 in precision metrics compared to the baseline model, while also reducing the size of model by 20%. In conclusion, the refined YOLOv8n algorithm presented in this study is an effective tool for detection in aerial images.