Wild-VLMDetect: A Vision-Language Framework for Injury Detection in WildlifeUsing YOLOv8, CLIP, and BLIP

Abstract: Wildlife injury detection is a crucial step in automating conservation efforts. This paper presents a hybrid vision-language model combining custom-trained YOLOv8 with CLIP and BLIP for injury recognition and description in wild animals. A dataset from Roboflow was used to train the YOLOv8 model for detection, while CLIP matched semantic prompts with cropped regions. Caption generation was handled by BLIP, enabling non-invasive, interpretable outputs. Quantitative evaluation included Top-k accuracy, mean similarity, prompt sensitivity, and caption fluency. The results demonstrate high accuracy for base classes and strong interpretability through visual-language integration. The proposed system offers a scalable and explainable framework for remote wildlife monitoring for injury detection. The proposed Wild-VLMDetect approach is a modular system that combines the strengths of YOLOv8 for object detection, CLIP for prompt-based semantic alignment, and BLIP for natural language captioning. Unlike WildCLIP, which focuses solely on attribute retrieval, Wild-VLMDetectintegrates detection, alignment, and captioning into a single pipeline. It also performs region-level detection and offers explain ability through descriptive outputs