Detection of Red Chili from Plant Images

Abstract

Manual identification of red chillies is often labour-intensive and inconsistent in large-scale harvesting, highlighting the need for automated, real-time agricultural systems. This study presents a lightweight red chilli detection system using a convolutional neural network (CNN) quantised to 8-bit TensorFlow Lite format and deployed on the PYNQ-Z2 FPGA board. The model classifies images into three categories: red chilli, plant without chilli, and unknown. Inference results are displayed using onboard LEDs and the PUTTY serial terminal, with LED 2 indicating red chilli, LED 1 for plant, and all LEDs off for unknown or low-confidence predictions. The system achieved a final training accuracy of 97.11%, a validation accuracy of 96.41%, and a test accuracy of 94.64%, demonstrating reliable classification performance with good generalisation and without signs of overfitting. Operating entirely offline without Jupyter or Ethernet, this low-power embedded AI implementation offers a practical, real-time alternative to manual chilli detection for smart agriculture applications.