A Lightweight Framework for Protected Vegetable Disease Detection in Complex Scenes

The rapid development of computer vision technology has provided new technical support for smart agriculture. Vegetable diseases represent a significant threat to agricultural production, with severity that cannot be ignored. However, through scientifically effective prevention and control measures, these negative impacts can be significantly mitigated. Intelligent disease detection systems, as advanced methods replacing traditional manual inspection, have become important means for developing smart agriculture and improving the efficiency of vegetable production management. Nevertheless, traditional manual detection is not only time-consuming and labor-intensive but also faces accuracy limitations, while existing computer vision detection methods still encounter a series of challenges when confronting complex backgrounds, diverse disease manifestations, and varying degrees of occlusion in real cultivation environments, including insufficient anti-interference capabilities, limited detection precision, and suboptimal real-time performance. This research addresses the practical challenges of limited data acquisition and sample scarcity for protected vegetable diseases by proposing an innovative strategy that implements differentiated data augmentation technique combinations for different categories of samples, significantly enhancing the model's resistance to environmental interference. Based on the integrated concepts of machine vision and deep learning, we developed a lightweight vegetable disease detection network named VegetableDet. This network innovatively combines Deformable Attention Transformer (DAT) with YOLOv8n backbone architecture, enhancing perception capabilities for long-range feature dependencies. Simultaneously, a Channel-Spatial Adaptive Attention Mechanism (CSAAM) is integrated into the Neck network, achieving precise localization and enhancement of key features. To address the issue of low model convergence efficiency, we further designed a hierarchical progressive transfer learning training strategy, effectively accelerating the model adaptation process and improving detection accuracy. Experimental evaluation demonstrates that on our custom comprehensive protected vegetable disease dataset, the VegetableDet model exhibits excellent performance in detecting 30 diseases and healthy samples across 5 vegetable types, with precision (P), recall (R), and average precision (AP) all exceeding 90%, and an overall mean Average Precision (mAP) reaching 94.31%. The model demonstrates powerful adaptability under complex environmental conditions, providing reliable technical support for real-time monitoring and precise prevention and control of protected vegetable diseases, with broad application prospects.