Study On Deep Learning Applied To Dog Face Identification,Detection And Generation

Despite the help of numerous pet welfare societies,to find a lost dog in the streets is often a difficult task for its owner.A free-ranging dog might also represent a potential danger for humans.It could carry diseases or become aggressive when wandering in pack defending a territory.Dog identification can help us take targeted protection or quarantine measures.Current identification techniques,as tags,tattoos and collars,lack of reliability or might be harmful for the animal.The development in computer vision field and,more specifically,in biometric identification could help to solve such difficulties.Cameras are cheap,efficient and now accurate ways to capture and identify objects thanks to deep learning,a recent machine learning technique.Deep learning has recently enabled the creation of reliable human face identification tools.Animal identification field in general could largely improve by using the advantages of deep learning.In this thesis,we thus present a deep learning approach toward dog identification.In addition to the aforementioned reasons,the choice of this specific specie has also been made to reduce the field of research and because the internet contains a lot of dog images needed for a deep learning method to properly work.In the first chapter,we present the related work on animal and dog identification and then develop on the general background of deep learning.In the second chapter,we create and pre-process a dataset of dog face images and develop several identification methods that can be trained on this dataset.We obtain an accuracy of 91.7%on the verification dataset with our newly developed adaptive quadruplet loss.In the following chapter we attempt to improve the previous datapreprocessing by creating an automatic face detection and alignment tool also using deep learning methods.We first enhance the dataset with face alignment thanks to 3D masks automatically applied on dog faces.We then implement a small deep neural network,inspired from the last development in object detection techniques,and apply it on the previous dataset.The network finally manages to learn how to detect dog faces in the wild.The final chapter of this thesis exposes several deep learning methods to further improve the drawbacks of our dataset,as blurry pictures or too few images per dogs.The main idea is inspired by the so-called Generative Adversarial Nets.We explain the different existing difficulties to train such networks and how we tackle them.We finally present our results which are still in need of improvement to fix our initial problem.This thesis thus contains a general pipeline of deep learning methods to automatically detect and identify dog faces in the wild.We hope that all these methods could help to further develop the domain of dog identification and more generally animal identification.