Contribuții la optimizarea și aplicațiile modelului deep convolutional neural networks / Deep convolutional neural networks

The purpose of this thesis is a set of contributions to optimization and applications of Deep Convolutional Neural Network (DCNN) model. This purpose is aligned to a growing global interest for the development of systems able to understand and learn from images in order to respond in case of some triggers or for initiating actions that enhance the actual context. A first application has been the classification of the drunkenness state and has been done from thermic infrared images of different faces using an ensemble of two DCNNs. We have explored the detection of the drunkenness state independent of the subject. In two different applications, we have classified the hyperspectral pixels from Pavia University dataset with custom DCNN architectures. For this objective we have compared multiple optimization algorithms, multiple loss functions and techniques as early stopping of the training simulation. In one of these last two mentioned applications, an original method (Generative Adversarial Networks, GAN) has been used for the creation of additional artificial training elements. The DCNN performance has been maximized in an application of credit requests classification, through training efficiency increase by optimization algorithms. The algorithms used for the optimum selection of training hyperparameters and for selection of some DCNN architecture elements are: Gradient Boosted Regression Trees, Decision Trees, Uniform Sampling, Random Search, Gradient Descent and Bayes Optimization. A world class DCNN architecture (YOLOv5) has been used for the context recognition in case of cars navigating urban areas. The DCNN hyperparameters have been selected in this case through a genetic algorithm driven by a configurable fitness function.