An Effective Agricultural Produce Object Detection System: YOLOv5-Based Classification

Automated object detection (OD) in real-time agricultural applications poses unique challenges due to uneven lighting conditions, object similarities, and cluttered environments. In comparison to standard object detection tasks, agricultural OD requires models that are both efficient and lightweight, enabling deployment on resource-constrained devices. This study introduces a novel agricultural produce detection system based on the YOLOv5 architecture, specifically designed to classify "Green Chilli," "Red Chilli," and "Eggplant." The YOLOv5 model was customized to handle these specific crops and optimized for deployment on edge devices by converting the model to TensorFlow Lite (TFLite).To enhance detection accuracy, advanced techniques such as data augmentation and hyperparameter tuning were applied. The TFLite model was optimized to run efficiently on IoT devices and low-powered CPUs, same as the Jetson Nano and Raspberry Pi, ensuring smooth real-time inference without the need for cloud-based computation. The SIoU loss function was utilized for precise localization, especially in densely packed images. The proposed system demonstrates high accuracy, precision, and recall, achieving a mAP) of 96% on a proprietary dataset of 600 images. The system's lightweight nature and real-time detection capabilities make it an effective tool for automating produce sorting and quality assessment in agricultural and food processing industries.