Classification of cashew kernel using deep learning

Cashew kernel grading is essential for determining the commercial value of kernels based on shape, size, and texture. Traditional manual methods are prone to inconsistencies due to human error, fatigue, and variable lighting, which reduce both quality and pricing accuracy. This paper presents a real-time automated cashew kernel grading system leveraging computer vision and deep learning. A YOLOv5s model, trained on a custom dataset comprising three commercial grades (W180, W300 and W500), was deployed on a Raspberry Pi. The system captures video input via a top-mounted webcam over a conveyor belt, performs kernel classification, and transmits the results to an Arduino Uno through UART. The Arduino controls mechanical flaps via L298N drivers to physically sort kernels. The system achieved over 93% classification accuracy, with physical sorting accuracy ranging from 94–96% across all classes. The proposed hardware-software co-design enables scalable, low-cost deployment in industrial environments, significantly improving throughput and reducing labor dependency.