Classification of five different varieties of walnut with deep learning models

Walnuts are an important agricultural product in terms of both nutritional value and economic return. Turkey ranks among the world's leading countries in walnut production, and many local and foreign walnut varieties with different morphological characteristics are cultivated. However, the correct classification of this diversity is critical in terms of producer selection, productivity, and quality control. Traditional classification methods are largely based on expert opinion and are prone to human error. Therefore, the automatic classification of walnuts through automated systems ensures accuracy, speed, and efficiency in agricultural production processes. In recent years, deep learning-based approaches have emerged as significant opportunities in this field and as alternatives to traditional methods. This thesis aims to classify different walnut varieties using deep learning models supported by image processing techniques. Within the scope of the study, a unique image dataset was created using samples from five different walnut varieties obtained from local producers in Sorhun Village, Niksar District, Tokat Province. This dataset was created by obtaining images of each walnut from a total of 144 different angles, both horizontally and vertically, at 5° intervals using a specially developed device. The data was processed using the Python programming language and the OpenCV library to perform pre-processing steps such as image resizing, contour cropping, and background cleaning, and was randomly divided into 50% training, 25% validation, and 25% testing. In the study, the performance of transfer learning-based models such as ResNet50, Xception, InceptionV3, MobileNet, and DenseNet121 was compared with the proposed Convolutional Neural Network (CNN) architecture. As a result of the model comparisons, it was observed that the proposed ESA architecture demonstrated remarkable performance with its low number of parameters, short training time, and high test accuracy of 99.78%. Furthermore, when confusion matrices, accuracy-loss graphs, and evaluation metrics were examined, it was noted that the proposed ESA architecture yielded more stable results.