基于改进YOLOv10n的葡萄叶片病害检测算法的研究

Research on Grape Leaf Disease Detection Algorithm Based on Improved YOLOv10n

摘要: 葡萄种植过程中，葡萄叶片病害在田间环境中一般会存在目标小、背景复杂、遮挡多等特征，并且现有的检测模型大多存在着参数量大、边缘部署难等问题。本文以YOLOv10n为基础，在此基础上提出一种轻量化的改进模型YOLOv10n-SH。该模型先用HGStem代替原来的浅层特征提取结构来减小高分辨率阶段的计算负担并且保留细节信息，然后加入由HGBlock和深度可分离卷积组成的特征提取模块，在保证参数量的同时提高主干网络的表达能力，同时在主干网络的三个特征层引入层次自适应融合模块HAFB，对C2f主路特征与HGBlock分支特征进行自适应对齐与加权融合，替代原有的拼接融合方式，从而增强小目标病斑的检测能力。实验结果表明，在自建葡萄叶片病害数据集上，与基线模型相比，YOLOv10n-SH的mAP@0.5达到97.7%，提高了2个百分点，参数量由2.3 M降至1.50 M，计算量由6.7 GFLOPs降至5.9 GFLOPs，具有一定的应用价值。

Abstract: During grape cultivation, leaf diseases generally exhibit characteristics such as small targets, complex backgrounds, and frequent occlusions in the field environment. Existing detection models often suffer from large parameter sizes and difficulty in edge deployment. Based on YOLOv10n, this paper proposes a lightweight improved model, YOLOv10n-SH. The model replaces the original shallow feature extraction structure with HGStem to reduce computational load at the high-resolution stage while retaining detailed information. It then incorporates a feature extraction module composed of HGBlock and depthwise separable convolutions, enhancing the backbone network’s expressive capability while maintaining the parameter count. Additionally, a hierarchical adaptive fusion module (HAFB) is introduced in three feature layers of the backbone network to adaptively align and weight the C2f main path features with the HGBlock branch features, replacing the previous concatenation fusion approach, thereby enhancing the detection capability for small disease spots. Experimental results on a self-built grape leaf disease dataset indicate that compared to the baseline model, YOLOv10n-SH achieves an mAP@0.5 of 97.7%, an increase of 2 percentage points, reduces the parameter count from 2.3 M to 1.50 M, and the computational load from 6.7 GFLOPs to 5.9 GFLOPs, demonstrating certain practical value.