A Hybrid Approach to Fine-Grained Butterfly and Moth Classification Using Deep Features and Rhombus-Based HOG Descriptor

Butterfly and moth species are crucial for ecosystems as pollinators, pestsand biodiversity indicators, therefore necessitating their precise automated classification for extensive monitoring, conservation initiatives, and agricultural pest control. Nonetheless, considerable obstacles emerge from interand intra-species variety in wing coloration, patterns, posture, and the effects of lighting and background circumstances on pictures. This study presents a comprehensive framework that enhances feature representation via a dual-phase methodology. Initially, pictures undergo preprocessing by Contrast-Limited Adaptive Histogram Equalization (CLAHE) to augment distinguishing features. Subsequently, elevated semantic features are derived using a ResNet50 backbone pre-trained on ImageNet, with a baseline accuracy of 92%. A unique Corner Rhombus Shape HOG (CRSHOG) descriptor is suggested to accurately capture detailed geometric and textural wing properties, utilizing rhombus-based grid sampling and gradient orientation encoding. These complementarydeep and handcrafted features are carefully integrated to form a hybrid representation, improving resilience to cluttered backdrops and position changes. The integrated feature set is assessed using several classifiers, with an Ensemble Subspace KNN modelattaining thegreatest classification accuracy of 94.6% on the Butterfly and Moth Image dataset, exceeding traditional CNN (Convolutional Neural Network)-only and HOG-based methods. These findings highlight the benefits of combining domain-specific shape descriptors with deep-learning features to enhance fine-grained insect categorization. Moreover, depending exclusively on standard RGB photos facilitates practical implementation on mobile and aerial platforms for real-time biodiversity surveillance and pest management.Future endeavors will concentrate on expanding this hybrid feature technique to encompass live video tracking and open-set species detection in uncontrolled settings.