YOLOv8-DDS: A lightweight model based on pruning and distillation for early detection of root mold in barley seedling

Abstract

• A one-stage detection model named YOLOv8-DDS was proposed to enable precise identification of minute mold-infected regions on plant roots. • The model underwent layer-adaptive magnitude-based pruning for parameter quantization, synergistically integrated with channel-wise knowledge distillation to achieve lightweight architecture while maintaining detection efficacy. • The optimized model achieved a detection latency of 57.6 ms with TensorRT acceleration, representing a 25.4 % improvement over the baseline model, which fulfills the real-time monitoring requirements in agricultural operational scenarios. • Integration of lightweight detection model on a rotatable vertical seedling cultivation rack. Root mold proliferation presents a significant challenge in the industrial production of hydroponic barley seedlings. The small size, inconspicuous coloration, and indiscernible image of early mold regions pose new demands on detection accuracy. This study constructed a dataset of root mold in barley seedlings throughout their growth cycle and proposed the YOLOv8n-DDS detection model to integrate a lightweight detection model into a three-dimensional cyclic cultivation system. The model incorporates the dynamic sample (DySample) operator, combines deformable ConvNets v2 (DCNv2) with C2f, and reconstructs the detection head using seam carving (SEAM) technology, which enhances its capability to extract multi-scale, minute features of early-stage root mold in barley. To improve the model’s performance on edge-embedded devices, this study employed layer-wise adaptive magnitude pruning and channel-wise knowledge distillation methods, thereby significantly reducing the model’s parameter count and computational load. The pruned and distilled model was subsequently deployed on the Jetson Nano platform for validation. Results indicate that the YOLOv8n-DDS model outperformed the baseline model in terms of precision, recall, and mAP50 by 2.4 %, 5.6 %, and 2.2 %, respectively. The parameter count was reduced by 23.8 %, and the computational complexity (Giga floating-point operators per second) was optimized by 14.8 %. Additionally, the detection latency on resource-constrained embedded devices was further reduced by 25.8 % with TensorRT acceleration. The proposed root mold detection model is lightweight and contributes to the intelligent and technological integration of the industrial production process for high-quality barley seedling forage.

Topics & Concepts