TMSDDET: A Lightweight Model for Tomato Maturity and Defect Detection on Edge Devices

Highlights TMSDDet enables accurate detection of tomato maturity and surface defects. The model achieved state-of-the-art performance with minimal computational cost. TMSDDet outperformed YOLOv6/8/11 in accuracy and speed. Model supports real-time inference on Jetson Orin Nano. Abstract. Accurate evaluation of tomato ripeness and surface quality is essential for intelligent grading and automated sorting in smart agriculture. However, traditional manual inspection methods are often inefficient and inconsistent, limiting their use in large-scale applications. To address these challenges, this article proposed TMSDDet (Tomato Maturity Surface Defect Detection) network, a lightweight detection model based on the YOLO11n framework, specifically designed for tomato maturity classification and defect detection. The model incorporates three optimized modules—ADown (a dual-path downsampling module), Slim-Neck (a lightweight multi-scale fusion structure), and Efficient-Head (an efficient decoupled detection head)—to achieve a strong trade-off between accuracy and computational efficiency. A multi-label tomato image dataset was constructed, including diverse ripeness levels and defect types under varying backgrounds. Experimental results demonstrated that TMSDDet achieved 80.4% mAP@0.5:0.95 with only 1.9M parameters, 3.9 GFLOPs, and an inference time of 6.2 ms per frame (RTX 4060). Compared to mainstream models such as YOLOv6, YOLOv8, and RT-DETR, TMSDDet delivered competitive performance on our test set while maintaining a compact model size of just 4.0 MB. Moreover, the model was successfully deployed on the Jetson Orin Nano without inference acceleration, achieving stable real-time performance (inference time: 31.0 ms). These results indicated that TMSDDet was an efficient, robust, and deployable solution for tomato freshness and defect detection, offering strong potential for practical deployment in resource-constrained agricultural environments. Keywords: Deep learning, Edge deployment, Lightweight model, Real-time detection, Tomato quality grading.