Deep learning based fruit separation system

With the developments in technology, computer vision systems have started to be used in many areas. Developing computer vision systems are also used in various ways in the field of agriculture. Fruits and vegetables go through many stages before they take their place on the shelves of markets and markets. One of the most important of these stages is to determine whether the fruits and vegetables are deformed and rotten, especially after the harvest. Today, after harvest, fruits and vegetables are separated by hand as rotten, damaged or intact. Manually sorting fruits and vegetables takes a long time and increases costs. It is possible to benefit from the developments in computer vision systems at the stage of classification of fruits and vegetables. Especially the high accuracy rates of deep learning models in image processing will be of great benefit in correctly automatically classifying and sorting fruits and vegetables as defective, rotten or healthy. By utilizing the automatic sorting of fruits and vegetables from computer vision systems, both time saving and cost reduction can be achieved. In this thesis, a deep learning-based parsing system was designed. A belted mechanical system that can automatically sort fruits and vegetables has been designed. It is designed to separate the fruits and vegetables moving over the belts into 3 classes according to their defective, rotten and sound condition. The system consists of two parts as real-time object detection and mechanical part with deep learning models. First of all, although there are different approaches for real-time object detection, since the YOLO algorithm is faster for real-time object detection, 6 different architectures of the YOLO algorithm are examined in this study. Classification of orange fruit as defective, rotten and intact was carried out to establish and test the system. A new dataset consisting of a total of 11238 orange images was created for the three classes. In the second part, the minimum hardware requirements of the system are determined. In the tests carried out on the prototype, up to 89.5% accuracy was obtained in the YOLOv5 model for 210 oranges.