AMET-YOLO: A Pear Leaf Disease and Pest Detection Model Integrating Multi-Strategy Feature Enhancement and Task Alignment

Accurate and robust detection of pear leaf diseases and pests is important for early warning in orchards, precision control, and intelligent disease and pest management. In natural environments, this detection task is still difficult because target scales vary greatly, lesion boundaries are often blurred, disease symptoms may look similar, small-target features are usually weak, and cluttered backgrounds can reduce detection performance. We develop AMET-YOLO as a task-specific extension of YOLOv11n. ADown is a downsampling operator borrowed from YOLOv9 that replaces selected stride-convolution layers and thereby preserves local details during scale reduction. The Memory-guided Sparse Expert Compensation Module (MSECM) is a feature-enhancement design which couples learnable memory retrieval with four dilation-specific experts and input-dependent Top-2 routing. In the neck, the Efficient Multi-scale Aggregation Fusion module (EMAFuse) substitutes four concatenation nodes and performs fixed-width channel alignment, element-wise aggregation, and lightweight depthwise–pointwise mixing. The Task-Aligned Detection Head (TAHead) is a modified decoupled head that retains the YOLOv11n prediction structure while it routes localization responses through a one-way gating path to modulate intermediate classification features. Experiments on the six-class PearLeaf-DP6 dataset show that AMET-YOLO achieves a precision of 88.5 ± 1.3%, a recall of 83.1 ± 0.9%, an mAP@50 of 88.7 ± 0.7%, and an mAP@50:95 of 51.8 ± 0.5%. Compared with the YOLOv11n baseline, AMET-YOLO improves these four metrics by 4.8, 3.9, 3.3, and 2.3 percentage points, respectively. Experiments on a second public tea leaf disease dataset further show that the proposed model stays effective when it is trained and evaluated independently on a different crop disease detection task. AMET-YOLO is a model with 6.32 M parameters and 10.2 GFLOPs, which constitutes a moderate rise in complexity relative to YOLOv11n while it remains far more compact than RT-DETR-ResNet50. These workstation-based results position AMET-YOLO as an accuracy-oriented image-based detector under the evaluated protocol, and real-time deployment on resource-limited orchard devices is left for future validation.