Fruit Detection and Classification using YOLO Models

Abstract

Computer Vision and Deep Learning techniques have become an advent in multiple domains like healthcare, Technology, as well as Agriculture . Computer vision techniques like object detection are being widely used in agriculture to reduce to efforts required and make agriculture a little more efficient for the farmers. The applications of deep learning in agriculture include leaf disease detection and weather forecasting, and the most advent applications include object detection to detect fruits, and vegetables which can be ensembled with robotics for automated yield production and harvesting. The proposed article describes one such application of fruit detection using various YOLO (You Only Look Once) models. The study encompasses four fruit classes namely Chiku, Mango, Mosambi, and Tomato. Models of Yolo V3, Yolo V4, and Yolo V8 were trained on a customized dataset collected from Indian farms and fruit gardens. The real time images images were collected, pre-processed, and annotated using online labeling tools. A total of 1200 images were used as a part of the complete training process. Basic preprocessing was performed on these images and possible inbuilt augmentation techniques supported by the above-mentioned models were used.Training is applied on custom dataset for all classes. In this experiment we have received the F1 score for YOLOv3(Chiku-82%.Mamgo-91%,Mosambi-87%,,Tomato-77%),YOLOv4(Chiku-89%.Mamgo-98%,Mosambi-95%,,Tomato-91%) and YOLOV8 (Chiku-90%.Mamgo-75%,Mosambi-82%,,Tomato-84%)models. In these models YOLOv4 with two layers gives the highest accuracy for all the classes.