Detection of One-Horned Rhino in Thermal Images Using Deep Model Architectures

Animal detection and surveillance have seen an upsurge in the research society due to the life endangerment scenario of animals on a large scale. While a lot of work has been done in this domain on day light images, surveillance during night time has an inherent drawback due to poor image quality issues like bad contrast, lousy illumination, etc. This work proposes to employ a single stage detector, namely YOLOv5, for the detection of the one-horned rhino in thermal images. This work also proposes a novel thermal dataset of the “one-horned” rhino. The annotation of the images was done manually for the purpose of using it in deep learning techniques. The automatic rhino detection system proposed herein uses a few altered backbone architectures like Mobile-Net, Shuffle-Net, EfficientNet, VGG-Net, and RepVGG-Net. Also, the ensemble operation of two models has been done to enhance the performance of the model. The performance of all the detector models was analyzed using the mAP and F1 score evaluation metric. It is observed that the ensemble operation improves the detection accuracy for thermal images.