Nextv2-DETR: lightweight real-time classification model of potatoes based on improved RT-DETR for mobile deployment

To address the challenge of quickly and accurately identifying and localizing potatoes in a complex production environment, this study proposes a lightweight potato classification algorithm, Nextv2-DETR, with enhanced feature extraction capabilities. The backbone of the model employs a lightweight ConNextv2 and incorporates DSC to reduce the number of parameters and realize potato feature extraction. The C2f module was enhanced with the BiFormer_attention mechanism, and a lightweight feature fusion network was designed. These modifications improved the extraction of potato features while maintaining a lightweight architecture. The average precision of the network saw a significant rise, climbing from 93.3 % to 96.3 % after the refinements were implemented. Concurrently, the model’s size, the floating-point operations required for processing a single image, and the detection timeframe were minimized to 43.8 %, 31.1 %, and 94.4 % of what they were initially. The mainstream target detection algorithm was employed for training on the dataset, and the training results were compared with the outcomes achieved by the algorithm proposed in this study. Compared to leading detection algorithms such as FAST-RCNN, YOLOv7x, YOLOXs, and lightweight models like YOLOv5s and YOLOv8s, the proposed method achieves superior accuracy with substantially fewer parameters. Operating at 70 FPS, the model provides a robust solution for integrating visual robotics into potato harvesting, demonstrating the potential to enhance agricultural productivity and efficiency.