Classification of walnut varieties using deep learning techniques

The walnut tree has historically stood out as an economically valuable tree species due to its edible fruit and high-quality wood. Walnuts are recognized as a significant component of a healthy diet due to their richness in Omega-3 fatty acids, antioxidants, protein, and fiber. Furthermore, walnuts contribute to brain health and heart health, bolstering the immune system and reducing inflammation. In the culinary world, walnuts hold a notable place, being used in various dishes and desserts across different cuisines. In the field of machine learning, deep learning has gained prominence, employing artificial neural networks to identify, classify, or predict complex structures and patterns within large datasets. Comprising multiple layers within the neural network architecture, deep learning algorithms process data to form high-level features or abstract concepts, combining information in more complex ways using non-linear activation functions. In my thesis, deep learning methods have been employed to classify walnut species with high nutritional value, aiming to categorize different walnut species based on their nutritional properties by analyzing their characteristics and patterns. In this thesis, deep learning-based classification was performed using images of walnut varieties Chandler, Fernor, Hekimhan, Kaman-1, Maraş-18, and Oğuzlar-77. A total of 2780 images were obtained, with 408 images of Chandler, 459 of Fernor, 485 of Hekimhan, 425 of Kaman-1, 393 of Maraş-18, and 610 of Oğuzlar-77. After preprocessing these images, the pre-trained models DenseNet121, NASNetMobile, and MobileNet were applied for classification. The dataset was divided into 80% training and 20% test data. As a result of the classification operations, 80,25%; 65,53% and 83,66% success rates were obtained, respectively. The highest classification success was achieved with the MobileNet model. However, parameter optimization was performed in order to increase the classification success. As a result, after optimization, the classification success of the pre-trained DenseNet121, NASNetMobile and MobileNet models was increased to 90,66%; 87,61% and 92,46%, respectively. Other detailed results obtained are presented in detail in the relevant sections of the thesis.