Integrated Deep Learning Framework for Animal Species Identification Using YOLOv5 and AOD-Net

Abstract

Encounters between humans and wildlife, particularly near roadways and urban peripheries, present notable dangers to both human safety and wildlife conservation. The development of automated systems for identifying animal species offers a promising strategy by providing early warnings to prevent vehicle-animal collisions and enhance ecological monitoring. However, achieving high accuracy in diverse environmental conditions such as haze, low light, partial obstructions, and complex natural backgrounds remains a significant challenge. This paper presents a robust, software-based framework that integrates YOLOv5, a state-of-the-art object detection model, with AOD-Net, an advanced image enhancement network specifically designed for haze removal and visibility improvement. In this hybrid approach, AOD-Net serves as a pre-processing module that enhances image quality by removing atmospheric distortions and improving feature clarity before the refined images are processed by YOLOv5 for object detection. The system was trained and evaluated on a diverse, labeled animal detection dataset containing multiple species across various environmental scenarios. Through comprehensive experimentation, the proposed model demonstrated significant performance gains over traditional YOLOv5 pipelines. It achieved a mean Average Precision (mAP) of 97.2%, with a precision of 95.26%, recall of 95.07%, and an F1 score of 95.13%. These results indicate superior detection accuracy and strong generalization across visually challenging conditions. The integration of AOD-Net significantly enhances the visual input quality, thereby improving the reliability of downstream object detection. This hybrid architecture offers a scalable and efficient solution for real-time wildlife monitoring and road safety systems, ultimately contributing to the mitigation of human-wildlife conflicts and the advancement of conservation efforts using AI-driven vision technologies.