Classification of hazelnut fruit varieties using deep learning method

Hazelnut is a type of nut that belongs to the Corylaceae family and is grown intensively worldwide, especially in Turkey. Turkey, with its provinces located in the Black Sea Region, is a leading center in hazelnut production and plays a significant export role worldwide. The examination of acres of hazelnut gardens by experts requires long periods of time and leads to high costs. In this study, it is aimed to classify eight of the hazelnut species commonly grown in Turkey using pre-trained networks. Within the scope of the study, images of hazelnut varieties were obtained through a computer vision system. The dataset we created includes 150 palaz, 320 mountain, 380 Foşa, 230 tombul, 380 honeysuckle, 320 çakdak, 420 black hazelnut, 390 spike, a total of 2590 images. These images were classified as hazelnut images with pre-trained models, DenseNet121, Inception v3, NasNet, EfficeNet, AlexNet and Xception. The dataset was divided into 80% training and 20% test sections. As a result of the classifications; 95,17%, 93,44%, 80,07%, 81,17%, 79,73% and 83.20% classification success was obtained, respectively. The highest classification success was obtained with the DenseNet121 model. Various confusion matrices and performance measurements were used to analyze the performances of the models in more detail. In the second stage, the images were multiplied using data augmentation techniques and subjected to re-classification. Pre-trained models applied to replicated data are; Densenet121, AlexNet, NasNet, Xception, Inception V3, VGG16, EfficeNet and VGG-19, respectively. The training success rates of these deep learning methods are; 97,5%, 91,16%, 81,61%, 82,77%, 95,45%, 92,95%, 78,66%, 82,95%. As a result of the data augmentation process, the classification success increased in some models and did not increase in others. The results obtained are given in detail in the relevant section of the thesis.