Improved YOLOv10 Deployment on Edge Devices for Tomato Grading

Abstract

Maintaining high tomato quality is crucial for consistent marketing standards and consumer satisfaction, necessitating the development of efficient automated tomato grading methods. Existing tomato grading systems are manual, labour-intensive, and time-consuming making them unsuitable for large-scale processing. Modern Deep Learning models are computationally expensive and not suitable for deployment on edges devices. Grading tomatoes is challenging because of the similarity in colour between some classes, such as unripe (green) and semi-ripe (yellow-green) tomatoes. This study proposes integrating a Convolutional Block Attention Module (CBAM) into YOLOv10 to improve feature extraction for enhanced accurate grading of tomatoes.  Furthermore, the proposed model is deployed on an NVIDIA Jetson Nano board to demonstrate its performance on a resource-constrained edge device. For this study, a dataset of 7058 tomato images was collected from tomato farms in Kenya. The dataset represents tomatoes in four stages of maturity: unripe, semi-ripe, ripe, and damaged and the model was trained to distinguish between these categories. On the Personal Computer, the model achieved a mean Average Precision (mAP) of 73%, an inference speed of 38.31 frames per second (FPS), and a model size of 5.812 MB. In the Jetson board, it achieved a mAP of 72.3% with an inference speed of 25 FPS using 640x640 image resolution, which is better compared to the competitors.