PugNet: A CNN Architecture to predict the Gender and Age of Blackbuck using Pugmarks

This study is on the Blackbuck identification from pugmarks, by applying advanced deep learning classification techniques to the BP dataset that contains 2 subclasses-age and gender that have been further divided into 3 categories implicitly. The initial approach for the problem was inspired from using the existing state of the art pretrained models to train the RGB images of the pugmarks of the Blackbuck that were the part of our BP dataset using Transfer Learning for multilabel classification. However, in this paper we have come up with a novel approach to challenge the accuracy and abilities of large pretrained architecture with a an optimized and improved Deep CNN architecture (PugNet) which takes binary masks generated from annotated images as inputs and produces results better than transfer learning approaches in age prediction and a better accuracy in case of gender predictions. Thus overall, the use binary masks helped us in reducing the environmental noise in each image and generate results which are better than the pretrained architectures. The proposed architecture also solves the problem of class imbalance which was observed to be the point of failure for the various pretrained architectures, The wildlife related datasets always face the issues of class imbalances due to presence of more female population in comparison to male population along with availability of samples older in age. Thus our novel PugNet solves the issue of class imbalances and also outperforms pretrained architecture in terms of predictions. The introduction of BP dataset will also in exploration of more complex image data with environmental noise acting as the major factor in performing the usual preprocessing techniques. The PugNet thus benchmarks the results on BP dataset and lays down a foundation for animal footprint identification problems and can be extended to various other species.