Simon Corbillé

Our goal is to design a tool for children's handwriting recognition and segmentation in order to accurately analyze the handwriting and provide immediate orthographic feedback to the child. The contributions of this thesis are based on the hybridization of deep learning models with models using explicit expert knowledge. The first contribution consists in integrating the writing dynamics of the online signal in a convolutional neural network for character recognition. The second contribution concerns the improvement of an existing word analysis system. This system uses a guidance mechanism based on the instruction and the words phonetically close to the instruction. It integrates the prediction of a Seq2Seq recognition model into the guidance system. The objective is to overcome the shortcomings of the guidance mechanism when the input words contain non-phonetic errors. The third contribution proposes a new recognition and segmentation framework. It is based on the combination of a model dedicated to recognition and a model dedicated to segmentation. The system also integrates the knowledge contained in the online signal in order to improve the accuracy of the segmentation. Finally, we have developed a rejection mechanism to improve the quality of the feedback given to the child. The results of the experiments demonstrate the interest and efficiency of these contributions.