Automated Pomegranate Disease Identification: A Deep Learning Pipeline with FANLM-FKM and HBOA

Abstract

Automated, Early and accurate identification of pomegranate Fruit diseases is vital for sustainable crop management and yield optimization. This research proposes a novel deep learning-based pipeline that integrates image preprocessing, instance segmentation, feature extraction, and hybrid classification for robust pomegranate disease diagnosis. The preprocessing stage employs a two-step enhancement technique combining Fuzzy Adaptive Non-Local Means with Fuzzy K-Means (FANLM-FKM) denoising followed by Contrast Limited Adaptive Histogram Equalization (CLAHE) for improving visual clarity while preserving structural details. Segmentation of diseased regions is performed using a fine-tuned Mask R-CNN model trained via the Detectron2 framework. Features are extracted from segmented images using the average pooling output of a pretrained ResNet-50 network. To reduce dimensionality and enhance discriminative power, the feature space is refined using a Hybrid Binary Optimization Algorithm (HBOA). The final classification is conducted using a custom-trained CNN optimized on the HBOA-selected features. Extensive evaluations reveal that the proposed CNN-HBOA classifier achieves high accuracy in distinguishing among four disease classes Anthracnose, Bacterial Blight, Cercospora, and Healthy FruitsF. The complete pipeline demonstrates strong generalization capabilities and holds promise for deployment in real-time agricultural diagnostic tools. This study outlines the methodology, implementation, and results, highlighting its potential agricultural impact.