【摘要】 在无人机检测领域，由于航拍任务对机载系统的严格限制，导致机载微处理器在存储和计算能力上存在显著的局限性。针对这一挑战，以YOLOv8作为基础模型，首先，借鉴轻量化卷积操作在减少模型参数数量方面的显著优势，将主干网络卷积块中的普通卷积替换为可分离卷积，并探索出一种结合轻量化卷积操作的结构块，保留C2f结构块丰富梯度流优势的同时缩小计算参数，两者交替构成新的主干网络结构，从而有效降低网络的规模和计算复杂度。此外，提出一种以Transformer解码器为基础优化后的轻量化全局感知组件，弥补轻量化网络设计导致的特征信息损失，实现特征的全局聚合。优化调整后的模型，参数量显著减少，检测精度有所提升，验证了模型设计的优越性。更多还原

【Abstract】 In the field of drone detection, the stringent constraints imposed on onboard systems by aerial photography tasks result in significant limitations in the storage and computational capabilities of onboard microprocessors. To address this challenge, this paper uses YOLOv8 as the base model. First, drawing on the significant advantage of lightweight convolution operations in reducing the number of model parameters, the ordinary convolution in the convolutional block of the backbone network is replaced with separable convolution, and a structural block combined with lightweight convolution operations is explored. While retaining the advantage of rich gradient flow of the C2f structural block, the computational parameters are reduced. The two alternately form a new backbone network structure, thereby effectively reducing the scale and computational complexity of the network. In addition, this paper proposes a lightweight global perception component optimized based on the Transformer decoder to compensate for the loss of feature information caused by the lightweight network design and realize global aggregation of features. The optimized and adjusted model significantly reduces the number of parameters while improving detection accuracy, thereby validating the superiority of the model design.更多还原