Animal facial detection for individual identification of animals using machine learning.docx

Abstract

In wildlife surveys and animal monitoring, cameras are frequently utilized. Depending on the trigger mechanism, there may be an accumulation of many pictures or movies. Studies have examined using deep learning algorithms to mechanically find animals in television camera pictures. This greatly minimize physical labor as well as rapidity up evaluation procedures. Few research, nevertheless, have compared and validated the usefulness of various object identification models in actual field monitoring circumstances. In order to conduct this investigation, we created an animal picture dataset from the AFRD dataset first. We also examined the credit presentation of trinity well-known thing discovery constructions as well as the efficacy of teaching shows using day-as well as-night data. For this research, feature extractors ResNet50, ResNet101, FCOS under the YOLOv5 series models, and Cascade R-CNN under HRNet32. The experimental results showed that the combined day-night training object detection models performed satisfactorily. Our models typically achieved 0.98 mean average precision and 88% accuracy in classifying animal videos and animal images, respectively. YOLOv5m completed the most accurate recognition in one stage. Ecologists can possibly swiftly and effectively extract information from vast amounts of photos with the use of AI technology, saving a lot of time.