Detection of olive fruit diseases with deep learning

This thesis presents a comprehensive analysis aimed at classifying two major diseases commonly observed in olive fruits physical damage caused by pests and deformations due to fungal infections through deep learning techniques applied to visual data. A total of 1,644 olive fruit samples were collected from orchards in the Bandırma district of Balıkesir province and categorized into three classes: "healthy," "pest-damaged," and "fungal-infected," based on expert evaluation. All samples were photographed under controlled lighting conditions using a specially designed imaging box. The resulting images were then processed using segmentation, thresholding, morphological operations, and object localization steps to prepare them for classification. In the modeling phase, nine different pre-trained CNN-based architectures (e.g., VGG16, ResNet50, InceptionV3, DenseNet121) and six attention-based Transformer models (ViT, BEiT, DeiT III, MobileViT, LeViT, MaxViT) were employed. Each model was trained with ten different optimization algorithms and comparatively evaluated in terms of classification performance. Additionally, existing MetaFormer-based models such as CaFormer, ConvFormer, and IdentityFormer were tested on the same dataset for systematic comparison. One of the key contributions of this study is the proposal of six novel hybrid architectures inspired by the MetaFormer framework. These architectures combine convolutional, random, identity, and Transformer blocks in various configurations, and were assessed using multi-dimensional criteria including classification accuracy, architectural complexity, and training time. Furthermore, a novel token mixer module named DeformationScore—which generates scores based on local texture density and edge sensitivity—was designed and integrated into the MetaFormer structure to form an entirely new model called DeformationScoreFormer. Model performances were thoroughly evaluated using metrics such as accuracy, F1-score, confusion matrix, and training-validation loss curves. The results demonstrated that both conventional CNN-based models and Transformer/MetaFormer-based architectures can accurately classify olive fruit diseases even under limited data conditions. Notably, the proposed hybrid MetaFormer models outperformed some existing models by achieving higher accuracy with lower computational cost in certain scenarios. This study highlights the effectiveness of deep learning approaches for disease detection and quality control in olive production and serves as a valuable resource for model selection and architectural design in agricultural image analysis.