基于曼哈顿距离注意力与门控双分支的桑叶多病害识别

Multi-disease Recognition of Mulberry Leaves Based on Manhattan Distance Attention with Gated Dual Branching

摘要:

桑叶病害导致桑叶质量下降、产量减少，进而影响蚕茧品质，制约桑蚕产业优质发展。目前，桑树种植存在病害识别智能化水平低、防治滞后等问题，且桑叶病斑特征复杂，细小病斑特征提取难度大。本文提出一种结合曼哈顿距离注意力增强与门控双分支结构的桑叶多病害复合识别算法YOLO v10s-MAD，对YOLO v10的骨干网络进行桑叶病虫害图像适配性优化，通过引入具有动态大核选择机制的网络结构（Large selective kernel network，LSKNet），增强对多尺度病斑特征捕捉能力，并减少冗余计算；同时，提出一个新的颈部网络MAD-Neck，通过引入基于曼哈顿距离的自注意力机制（Manhattan distance-based self-attention mechanism，MaSA），并结合Transformer架构的多尺度动态大核门控双分支模块（Multi-scale gated dual-branch module，MGDB），更好地对多种病害的细节特征识别；增加归一化沃瑟距离（Normalized Wasserstein distance，NWD）损失函数增强模型对小目标检测中的微小位置偏差的鲁棒性。不同模型性能试验结果表明，改进算法模型相较于原模型，mAP50和mAP50:95分别提升2.2、3.0个百分点，达到92.2%和76.8%，满足对桑叶病害进行目标检测的要求。

Abstract:

Mulberry leaf diseases cause significant degradation in leaf quality and yield reduction, adversely impacting silkworm cocoon production and constraining the high-quality development of the sericulture industry. Current challenges encompass low levels of intelligent disease identification, lagging prevention and control measures, and the inherent difficulty in extracting discriminative features from complex lesion morphologies, particularly for fine-scale disease spots. To address these limitations, a novel multi-disease recognition algorithm, YOLO v10s-MAD was proposed, integrating a Manhattan distance-based self-attention mechanism (MaSA) within a gated dual-branch structure. The backbone network of YOLO v10 was specifically optimized for mulberry disease imagery through the integration of a large selective kernel network (LSKNet). This module employed a dynamic large kernel selection mechanism to significantly enhance the model’s capability for capturing multi-scale lesion features while simultaneously mitigating computational redundancy. Furthermore, a newly designed neck network, MAD-Neck, was introduced. MAD-Neck incorporated the MaSA mechanism, which utilized Manhattan distance to compute attention weights more efficiently, focusing model capacity on salient pathological regions. It also integrated a multi-scale gated dual-branch module (MGDB), incorporating structural principles from the transformer architecture, to effectively fuse features across different scales and improve discrimination between subtle disease characteristics. To enhance robustness specifically for detecting small lesions, the normalized Wasserstein distance (NWD) loss function was adopted for bounding box regression, reducing sensitivity to minor localization deviations common with tiny targets. Comprehensive evaluations demonstrated that the enhanced model achieved 92.2% mAP50 and 76.8% mAP50:95, representing improvements of 2.2, 3.0 percentage points over the baseline, respectively, fulfilling practical deployment requirements for accurate mulberry disease identification.