Improved YOLOv11 Framework with Channel Attention for High-Precision Identification of Luffa Leaf Diseases

Early accurate identification of luffa crop diseases is essential for timely treatment and high-quality agricultural production. Currently, luffa disease diagnosis relies solely on manual field inspection, which is labor-intensive and error-prone, and lacks assessment uniformity. While deep learning-based diagnostic methods have emerged as a promising solution, existing image classification frameworks fail to achieve an optimal balance between identified accuracy and resource efficiency, making them incompatible with high-performance real-time monitoring on resource-limited outdoor edge devices. To address these issues, this paper proposes a lightweight, high-efficiency luffa leaf disease identifying framework based on the YOLOv11 architecture. Two structural optimizations are implemented to enhance the model’s comprehensive performance. First, the original backbone network is replaced with MobileNetV3-Small, and the down-sampling module is optimized to reduce model parameters and computational complexity while preserving critical spatial feature stability. Second, a Squeeze-and-Excitation (SE) module is integrated into the feature fusion neck to recalibrate channel-wise feature weights. This enables the network to focus on salient pathological features—including the characteristic yellow rings of Alternaria lesions and irregular textures of mosaic lesions—while suppressing background noise from soil and interfering vegetation. The YOLOv11 baseline is adopted for its superior feature pyramid network structure and optimized training convergence performance. Evaluated on the publicly available LuffaFolio dataset, the proposed model achieves a highly competitive efficiency balance compared with mainstream lightweight detectors such as YOLOv8n and YOLOv10n. Specifically, the optimized model attains an overall accuracy of 97.65%, maintaining a performance strictly comparable to the original YOLOv11n baseline, while reducing model parameters and computational overhead by 41.54%. It achieves a precision of 96.85% and a recall of 97.12%, with an ultra-small model size of only 3.1 MB. Experimental results demonstrate that the proposed framework serves as a promising proof-of-concept, showcasing a favorable accuracy-to-complexity trade-off for future real-time agricultural applications.