Deep Learning for Biodiversity: Multiclass Animal Image Recognition

Abstract

Abstract: Abstract- Human-Animal Conflict presents serious problems including threats to human life and resource loss, which calls for careful monitoring and preventive actions. This work aims to create an advanced animal detection and identification system leveraging motion-triggered cameras, sometimes referred to as camera traps, images obtained. Poor contrast between animals and cluttered backgrounds as well as high false-positive rates brought on by dynamic backgrounds cause existing techniques to suffer with low detection rates. We propose a two-stage approach to overcome these difficulties: first, using advanced image processing techniques to generate animal object region proposals; second, using artificial intelligence (AI) methods for detection and classification. The system validates region proposals and accurately identifies animals using eXtreme Gradient Boosting (XGBoost), Particle Swarm Optimisation (PSO), and CNN. The suggested system finds use in animal intrusion detection in industrial settings close to residential areas, agricultural fields, factories, and avoidance of animal-vehicle collisions. This method seeks to raise detection accuracy and dependability, so supporting more effective and safer methods of human-animal coexistence.