Automatic Hair Segmentation

Cilj ovog rada je izrada modela za klasifikaciju kose u stvarnom vremenu. Glavno sredstvo kreiranja dubokih neuronskih mreža bio je jezik Python te dodatak PyTorch. Navedena je teorijska pozadina modela DenseNet i Tiny DarkNet temeljenih na konvolucijskim neuronskim mrežama te pripadajući implementacijski detalji. Prikazani su podaci za trening kao i njihova priprema te utjecaj istih na robusnot podataka. Takod̄er napravljena je usporedba treniranih modela koji se razlikuju po hiperparametrima. Najbolji rezultati su pokazali kako je moguće ostvariti vrlo efikasno i brzo rješenje koje ispunjava uvjete pokretanja na mobilnim ured̄ajima.The aim of this paper is to create a model for hair classification in real time. The main tool for creating deep neural networks was the Python language and the PyTorch plugin. The theoretical background of the DenseNet and Tiny DarkNet models based on the convolution neural network were described as well as the related implementation details. The data is presented for training as well as their preparation and their influence on the robust of data. Also a comparison of trained models differing in hyperparameters were presented. The best results have shown that it is possible to achieve a very efficient and fast solution that meets mobile start-up conditions