Automated colour-based pepper sorting

Abstract

Bell pepper sorting is an important post-harvest process that influences product quality, market value and processing efficiency. Traditional manual sorting methods are labour-intensive, time consuming and susceptible to human error, while many automated sorting systems remain expensive and inaccessible to small-scale farmers. This project presents the design and implementation of an automated bell pepper colour sorting system based on computer vision and TinyML technologies. The proposed system consists of an ESP32-CAM module for image acquisition and colour classification, an ESP32 microcontroller for system coordination, an infrared sensor for pepper detection, a conveyor belt driven by stepper motors, and servo-actuated sorting mechanisms. An Edge Impulse TinyML model was trained using a dataset of 600 training images and 152 testing images across four classes, green, yellow, red and background. The trained model was deployed directly on the ESP32-CAM for real-time inference. Experimental results showed that the Edge Impulse model achieved 100% precision, recall and F1-score on the validation dataset. When integrated into the complete prototype, the system achieved an operational sorting accuracy of 99.6% under normal lighting conditions. The average inference time was approximately 150 ms per image, while the system processed one pepper every two seconds. The results demonstrate that low-cost embedded machine learning can provide accurate and reliable agricultural sorting without requiring high-performance computing hardware. The developed prototype offers a practical, affordable and scalable solution for automated pepper sorting and demonstrates the potential of TinyML-based computer vision systems in agricultural automation.