Coffee plant disease identification with an attentive multi-image segmentation framework (MISF) with CycleGAN

• A reliable usage of the Attentive Segmentation approach to find ROI. • A novel training process in two phases of the proposed transfer learning models with segmented images and without segmented images. • Understanding the impact of model parameters with various iterations. • Analyzing the importance of synthetic data generation for rare crops like Coffee. This research presents a novel approach to coffee plant disease identification by leveraging an attentive image to-image translation system based on CycleGAN. The proposed system enhances the diagnostic process by transforming images of healthy coffee plants into diverse representations of diseased states using the Multi-Image Segmentation Framework (MISF). Integrating an attention mechanism allows the model to focus selectively on pertinent regions, contributing to improved ac curacy in disease identification. Experimental evaluations on a comprehensive dataset demonstrate the approach's effectiveness, showcasing its potential as a robust and efficient tool for automated coffee plant disease diagnosis. Hence, the proposed framework implements a strong and reliable synthetic image-to-image data generator based on CycleGAN, referred to as Extensified CycleGAN, along with the automated classification model ResNet101. The proposed work incorporates the benefits of pre-trained models in developing an architecture for synthetic data generation and disease identification by addressing the data imbalance problem effectively. This research study introduces a novel attentive Multi-Image Segmentation Framework (MISF) combined with an enhanced CycleGAN architecture, tailored specifically for coffee plant disease diagnosis. The innovation lies in integrating region-focused segmentation prior to synthetic image generation, enabling the GAN to produce biologically meaningful and lesion-specific training data. Unlike conventional GAN-based augmentation approaches, this segmentation-driven pipeline emphasizes disease-affected regions, thus improving the discriminatory learning capability of CNN classifiers. Quantitatively, the ResNet101 model trained with the proposed augmented dataset achieved an overall classification accuracy of 98.38%, significantly outperforming its counterparts trained without segmentation, which recorded only 91.93% accuracy. Notably, precision and recall for individual diseases such as Rust and Phoma reached up to 98.9% and 98.1%, respectively. These results clearly validate the contribution of the segmentation-aware synthetic augmentation strategy, demonstrating its effectiveness in addressing data imbalance, enhancing classification accuracy, and ensuring better generalization in real-world agricultural scenarios. The Multi-Image Segmentation Framework (MISF) is a composite segmentation technique designed to accurately extract the region of interest (ROI) from coffee leaf images. It combines multiple classical methods—thresholding, K-means clustering, Watershed, GrabCut, and Mean Shift—to isolate disease-specific lesions while eliminating irrelevant background elements. This targeted segmentation ensures that the subsequent synthetic image generation and classification processes focus only on meaningful disease features, thereby enhancing model accuracy and generalization. In comparison to existing state-of-the-art methods that rely on generic GAN-based augmentation or shallow CNN classifiers, our approach integrates attentive segmentation with an enhanced CycleGAN to generate lesion-focused synthetic data. This targeted augmentation significantly improves feature learning for rare disease classes, leading to more balanced training. As a result, our ResNet101-based classifier achieved an overall accuracy of 98.38%, outperforming conventional models that typically range between 90–93%.