Modification of C-Grabcut for Segmentation and Classification of Coffee Leaf Diseases in Complex Backgrounds

Visual changes, including spots, discoloration, and deformation characterize coffee leaf diseases. In real-world image data, complex backgrounds present challenges for classification using deep learning models. Irrelevant objects, such as soil, other leaves, and miscellaneous items, can hinder the model's ability to accurately recognize disease patterns. Furthermore, the absence of effective segmentation techniques has resulted in low accuracy in previous studies. This work aims to address these limitations by enhancing the performance of the MobileNet-V2 model for coffee leaf disease classification. We applied a modified C-Grabcut segmentation technique to improve the isolation of diseased areas from complex backgrounds. The results demonstrate a significant performance improvement, achieving an Intersection over Union (IoU) of 0.8369 and an accuracy of 94.83%. These findings suggest that the modified MobileNet-V2 model, combined with the improved C-Grabcut segmentation, offers robust performance for in-field coffee leaf disease classification, striking a better balance between effectiveness and accuracy compared to previous studies.