An Investigation into Improved YOLOv8-based Target Detection Algorithms for UAV Aerial Imagery

• Published in: 2023 International Conference on Image Processing, Computer Vision and Machine Learning (ICICML) pp. 164 - 170
• Main Authors: Ming, Rui, Liu, Xinyu, Lin, Yuxuan, Gong, Qian
• Format: Conference Proceeding
• Language: English
• Published: IEEE 03.11.2023
• Subjects: Autonomous aerial vehicles; Computational modeling; Convolution; Multi-scale Convolutions; Object detection; Optimization; Routing; Target Detection; Target recognition; UAV Aerial Imagery; YOLOv8-based
• DOI: 10.1109/ICICML60161.2023.10424855

In UAV imagery, the intricate backgrounds combined with the high quantity and compact distribution of minute targets have consistently made target detection a formidable challenge in the realm of computer vision. This study introduces an enhancement over the YOLOv8 algorithm, wherein a sophisticated multi-scale convolutional layer, integrating depth-separable convolution, attention mechanisms, and multi-scale processing techniques, replaces the original model's convolution. Moreover, we introduce an attention mechanism for a Bi-Level Routing within the core component of the base model, and adjustments are made to the original model's loss function. Lastly, to confirm the viability of the enhanced model proposed in this paper, we conducted a validation of the metrics using publicly accessible datasets. The findings illustrate that the improved model outlined in this research substantially enhances target recognition accuracy in UAV images. Furthermore, the model exhibits superior performance in mitigating issues of duplicate detection and target omission.