ENHACLIP-SEARCH: CLIP-BASED VISION-LANGUAGE LEARNING FOR WILDLIFE CLASSIFICATION IN THE HIMALAYAN ECOSYSTEM

Abstract

Biodiversity loss and the need for efficient ecological monitoring have increased the demand for reliable and automated wildlife species recognition systems. However, conventional image classification methods often struggle in complex natural environments due to variations in background, illumination, pose, scale, and species appearance. To address these challenges, this study proposes EnhaCLIP-Search, a vision-language framework for fine-grained wildlife classification in natural scene images. The proposed approach is built upon the CLIP ViT-B/32 architecture, which integrates a Vision Transformer for extracting visual representations with a text encoder for generating semantic representations of species descriptions. Image and text embeddings are projected into a shared feature space and compared using cosine similarity to identify the most semantically compatible wildlife category. To improve robustness under real-world environmental conditions, eight data augmentation strategies incorporating variations in shape, color, and illumination are employed during model development. The proposed framework focuses on fine-grained visual characteristics, including bird plumage and mammalian fur patterns, to distinguish between 20 wildlife species native to the Himalayan region. Experimental evaluation on a self-curated dataset demonstrates that EnhaCLIP-Search achieves 95% accuracy, 96% precision, 94% recall, and a 95% F1-score, outperforming the evaluated baseline approaches. These results demonstrate the effectiveness of the proposed vision-language framework for robust wildlife species recognition and highlight its potential for automated biodiversity monitoring in complex natural environments.