Învățare cu marjă largă pentru analiza imaginilor / Large margin learning for image analysis

In the last decade, deep learning has developed rapidly due to the interest generated in the community which was based on achieved performance but also making use of the significant improvement of hardware resources. Thus, automatic learning algorithms have become increasingly attractive to researchers in numerous fields. The main goal of the thesis was to improve techniques based on convolutional networks. One of the impediments encountered very often in this field is represented by the small amount of data or non-compliant labels, which have a bad influence on approaches that use supervised learning. In an effort to combat these limitations, a number of new methods were proposed in this work. To take advantage of unlabeled data, semi-supervised learning techniques, transfer learning, and novel methods of augmentation and regularization were used. Another common issue in automatic learning is the overlap of data in the descriptor space, which negatively impacts the separation boundary and the network's predictions. In order to mitigate this effect, a new loss function was proposed to group the descriptive space more efficiently. The main topics addressed were the recognition of facial expressions and image retrieval. n the case of facial expressions, the emphasis centered on identifying the type of emotion, but experiments to identify facial movements were also conducted. The second area of research was based on obtaining embeddings based on convolutional networks trained with a new clustering loss function. The results for both fields proved satisfactory and confirmed the potential of the proposed methods.