A novel measurement method for detecting pepper diseases and pests in greenhouse environments based on improved YOLOv11

Abstract While protected cultivation in greenhouses enhances pepper yield, it exacerbates the proliferation of diseases and pests, rendering manual monitoring labor-intensive and inefficient. Automated detection in such environments, however, faces severe challenges arising from complex foliage occlusion, variable lighting conditions, and extreme scale variations among small pests and large lesions. Existing lightweight detectors often compromise accuracy for speed, failing to meet the demands of precision agriculture. To address these conflicts, this paper proposes PepperDiseaseDet, a high-precision, low-latency detection framework tailored for complex greenhouse scenarios. We introduce three novel, synergistically integrated modules: (1) a channel-gated feature aggregation module that reconstructs global spatial skeletons to robustly identify targets under severe occlusion; (2) an adaptive factorized convolution module that replaces redundant computations in the neck network, acting as a computational throttle to enhance inference speed without sacrificing feature integrity; and (3) a cross-scale adaptive fusion module that resolves semantic inconsistencies across resolutions, ensuring the precise concurrent detection of microscopic pests and expansive disease spots. Rigorous evaluations on a custom, challenging greenhouse dataset demonstrate that PepperDiseaseDet achieves an mAP@0.5 of 94.3%, a precision of 92.5%, and a recall of 96.0%, significantly outperforming the state-of-the-art YOLOv11n baseline by 4.1%, 2.8%, and 4.2% points, respectively. Crucially, the model achieves an inference speed of 7.0 ms on standard hardware, with reduced parameter count (2.4 M) and FLOPs (5.9 G). These results establish PepperDiseaseDet as an optimal solution balancing high accuracy and computational efficiency for real-time edge deployment in smart greenhouse operations.