CDMS-YOLO: An Enhanced YOLOv11 Model with Gated Channel Attention and Dynamic Multiscale Optimization for Coffee Leaf Disease and Pest Detection

Accurate and efficient detection of coffee leaf diseases and pests is essential for safeguarding coffee yield and quality. To address the challenges posed by small lesions, complex backgrounds, and highly variable symptom morphologies under field conditions, this study proposes CDMS-YOLO, a lightweight detection model based on YOLOv11n. CDMS-YOLO integrates the C3k2-Convolutional Gated Linear Unit (C3k2-CGLU) module, the DySample dynamic upsampling operator, the Mobile Inverted Bottleneck Convolution detection head (Detect_MBConv) detection head, and Scale-based Dynamic Loss (SD Loss) to enhance disease feature extraction, multiscale detail recovery, and bounding-box regression. Experiments on a coffee leaf disease and pest dataset with complex backgrounds show that CDMS-YOLO achieves a precision of 93.0%, a recall of 90.1%, and an mAP50 of 95.1%, representing improvements of 4.8, 2.1, and 3.0 percentage points, respectively, over YOLOv11n. Under a stricter leakage-controlled data split, CDMS-YOLO achieves an mAP50 of 93.6% and an mAP50-95 of 77.3%, outperforming YOLOv11n by 3.1 and 2.6 percentage points, respectively, indicating that the performance advantage is maintained under a more conservative evaluation protocol. The model contains only 2.84 M parameters and requires 7.2 GFLOPs. On the NVIDIA Jetson Orin Nano, CDMS-YOLO achieves an inference speed of 135.41 FPS, compared with 203.75 FPS for YOLOv11n. Thus, its frame rate is approximately 33.5% lower than that of YOLOv11n; nevertheless, it still satisfies real-time detection requirements. Overall, CDMS-YOLO provides an effective balance between detection accuracy, model complexity, and edge-deployment efficiency, offering a practical approach for intelligent detection and precision management of coffee leaf diseases and pests.