An Automated Boiled Arecanut Grading Using Deep Learning

Abstract

The commercial value of boiled arecanut is determined largely by its visual grade, and the process by which this grade is assigned continues, in most processing units, to depend on the manual judgement of experienced workers. Such an approach is slow, requires a substantial workforce, and is prone to inconsistency because human assessment varies with fatigue, lighting conditions, and individual perception. This paper presents an automated grading framework that applies deep learning and computer vision to classify boiled arecanut samples into four commercially recognized categories — Hasa, Bette, Rashidi, and Gorabalu — directly from digital images. A dataset of 1,792 images was collected under controlled illumination in Shivamogga district, Karnataka, and subjected to resizing, ImageNet-based normalization, and augmentation before being used to train three transfer-learning-based convolutional neural networks, namely AlexNet, GoogLeNet, and ResNet50. The three architectures were benchmarked against one another using accuracy, precision, recall, and F1-score, with AlexNet emerging as the strongest performer at a test accuracy of 99.15 percent. To make the system usable outside a research setting, a Streamlit-based web interface was built that allows a processing unit to upload an arecanut image and receive an immediate grade prediction along with a confidence score. Gradient-weighted Class Activation Mapping (Grad-CAM) was further incorporated to visualize the image regions that most influenced each prediction, giving operators a transparent view of the model's reasoning rather than a black-box output. Collectively, the results indicate that the proposed system is capable of replacing subjective manual grading with a fast, consistent, and interpretable alternative suited to real-world agricultural processing.