Abstract: Abstract: To address the degradation in cherry disease detection accuracy caused by leaf occlusion and background clutter in complex agricultural environments, an improved variant of YOLOv8n named YOLO-APA was proposed. Three key enhancements centered on feature extraction and fusion were introduced. An Adaptive Focus Attention (AFA) mechanism with channel?spatial co?optimization was designed to enhance the model’s ability to recognize occluded regions. A Cross?position Pyramid Module (CPPM) was constructed, which integrated positional encoding with an adaptive gating mechanism to better fuse multi?scale features; the adaptive gating was used to extract local detail features and global context information respectively, and feature complementarity was achieved through channel?wise concatenation. An Adaptive Grouped Feature Fusion (AGFF) module was employed to dynamically group and fuse multi?scale features, aiming to improve the multi?target discrimination capability under complex backgrounds. Experiments were conducted on a self?constructed cherry disease image dataset. The results showed that YOLO?APA achieved a precision of 61.2%, a recall of 61.8%, and an mAP?.? of 63.4%, which were improvements of 5.0, 11.7, and 14.6 percentage points, respectively, over those of the original YOLOv8n.

Key words: Keywords: cherry disease detection, object detection, attention mechanism, shot hole disease, brown rot, gummosis

摘要： 摘 要: 针对复杂农业场景下枝叶遮挡、背景干扰导致樱桃病害检测精度下降的问题，提出一种基于YOLOv8n改进的YOLO-APA模型。围绕特征提取与融合进行三方面改进：设计通道-空间协同优化的AFA自适应聚焦注意力机制，增强模型对遮挡区域的识别能力；构建CPMM跨位置多尺度特征提取模块，融合位置编码与自适应门控机制并使用更好融合多尺度特征、引入自适应门控分别提取局部细节特征与全局上下文信息，并通过通道拼接实现特征互补；使用AGFF自适应分组特征融合模块，对多尺度特征进行动态分组与融合，以提升模型在复杂背景下的多目标判别能力。实验基于自建樱桃病害图像数据集，结果显示，YOLO-APA模型的精确率达到61.2%，召回率为61.8%，平均精度mAP为63.4%，相比原始YOLOv8n分别提升9.0、11.8和14.6个百分点。

关键词: 关键词: 樱桃病害检测, 目标检测, 注意力机制, 穿孔病, 褐腐病, 流胶病