Authors

Abstract

The detection of apple leaf lesions in complex environments is hindered by several factors, such as the small size of lesion areas, variability in lighting conditions, and occlusions caused by overlapping leaves. These issues significantly limit the performance of existing detection models. Therefore, an enhanced detection algorithm for apple leaf lesions, termed PAME-YOLO, is proposed in this study, building upon the YOLOv8s framework. First, the main convolutional module is reconstructed using the Parallelized Patch-Aware Attention Module (PPA) while fusing Efficient Multi-Scale Attention (EMA). This effectively strengthens the model’s capacity to localize small target lesions in complicated environments. Second, an Attention-based Intra-scale Feature Interaction (AIFI) is introduced into the feature extraction network to replace the Spatial Pyramid Pooling-Fast (SPPF) module, which better captures the subtle features of apple leaf lesions. Next, the downsampling enhancement module is designed to mitigate information loss during the original downsampling process, which contributes to a significant improvement in detection precision. Finally, the Efficient Head is designed, a lightweight and efficient detection head that lowers parameter count and computational intricacy without sacrificing accuracy. Compared with YOLOv8s, the proposed model delivered a notable enhancement in performance, with precision (P) increasing by 0.8 points and recall (R) by 1.5 points. The mAP@0.5 achieved 91.4%, which is 1.5 percentage points higher than that of YOLOv8s. Meanwhile, the mAP@0.5:0.95 rose to 56.4%, reflecting an increase of 1.4 percentage points. The improved model realizes the accurate detection of apple leaves lesions in complicated surroundings, offering reliable technical assistance for disease prevention and contributing to the development of the apple industry.

Abstract in Chinese

在复杂环境中，苹果叶片病斑的检测受到诸多因素的影响，如病斑区域尺寸较小、光照条件变化以及叶片重叠造成的遮挡等问题。这些因素严重限制了现有检测模型的性能。为此，本文在YOLOv8s算法的基础上提出了一种称为PAME-YOLO的苹果叶片病斑检测算法。首先，本文使用并行补丁感知注意力模块同时结合高效多尺度注意力机制对主干卷积模块进行重构，这有效增强了模型在复杂环境中对小目标病斑的定位能力。其次，本文在特征提取网络中引入基于注意力的尺度内特征交互模块，来替换原有的快速空间金字塔池化模块，以更好地捕捉苹果叶片病斑的细微特征。随后，本文设计了新的下采样增强模块，以弥补原有下采样过程中的信息丢失，从而显著提高检测精度。最后，我们设计了一种轻量高效的检测头Efficient Head，该检测头能够在保持精度的同时降低模型参数和计算复杂度。与YOLOv8s相比，所提出的模型在性能上取得了显著提升，精确率（P）提高了0.8个百分点，召回率（R）提高了1.5个百分点，mAP@0.5达到了91.4%，比YOLOv8s高出了1.5个百分点，同时mAP@0.5:0.95达到了56.4%，提高了1.4个百分点。综上所述，改进后的模型实现了在复杂环境下对苹果叶片病斑的精准检测，为病害防控提供了可靠的技术支持，助力了苹果产业的可持续发展。