Abstract: Aiming at the problem of poor object detection effect caused by the degradation of UAV image quality in complex weather such as rain and fog, a target detection algorithm CGP-YOLO (context-guided and prompt-based YOLOv8) based on context guidance and prompt learning is proposed. A multi-task joint learning detection network is constructed to balance detection and image restoration tasks through a two-branch structure. A cross-layer attention-weighted denoising branch based on prompt learning is proposed to guide the network to reconstruct clear images using degradation prompts. The model backbone is designed with a context-based residual sampling module, and the convolutional attention mechanism is integrated into it, so that the local and global information of the target can be integrated. The separable large-core multi-scale feature extraction module is used to process the multi-scale features of the network. The special detection head of small object is introduced to enhance the detection accuracy of small object. The experimental results show that with only 60% of the parameters of the baseline model, the detection accuracy of the model gets an increase of 2.4 percentage points, and the average accuracy (mAP) is increased by 2.04 percentage points. The model detection effect is better than other classical models and has excellent performance.

Key words: multi-task learning, object detection, UAV image, complex weather, prompt learning, denoising model

摘要： 针对雨雾等复杂天气下无人机图像质量下降导致目标检测效果不佳的问题，提出基于上下文引导和提示学习的目标检测算法CGP-YOLO（context-guided and prompt-based YOLOv8）。构建一个多任务联合学习的检测网络，通过双分支结构达到平衡图像检测和恢复的任务。提出基于提示学习的跨层注意力加权图像去噪分支，指导网络利用退化提示重构清晰的图像；模型主干设计基于上下文的残差采样模块，集成卷积注意力机制，综合目标的局部和全局信息；采用可分离大核多尺度特征提取模块，处理网络多尺度特征；引入小目标的专用检测头，增强小目标的检测精度。实验结果表明，在参数量仅为基线模型60%的情况下，该模型的检测精度提高了2.4个百分点，平均精度（mAP）提高了2.04个百分点，模型检测效果优于其他经典模型，具备卓越的性能。

关键词: 多任务学习, 目标检测, 无人机图像, 复杂天气, 提示学习, 去噪模型