Klasifikace ručně psaného textu s využitím konvolučních neuronových sítí

Abstract

The work focuses on character recognition in images within the scope of machine vision. It describes basic techniques for image processing and manipulation from initial preprocessing to more complex image operations, such as classification using neural networks. The insights presented aim to provide readers with the necessary information to create an algorithm for detecting numerical sequences, which can be applied, for example, in postal services or banks. The main objective is the segmentation of numerical characters and the identification of their individual values. Convolutional neural networks are employed for character recognition, trained on both custom data and data available in MATLAB for comparison. Custom training data are obtained from photographs of handwritten digits from various individuals. An algorithm developed in MATLAB, along with additional functions utilizing pre-trained networks, is designed and implemented for processing and segmenting these numbers. A significant portion of the work deals with segmentation and other image processing procedures necessary for obtaining input data in the required format for the convolutional neural network. The results are presented in a graphical user interface application specifically designed for recognizing numerical sequences on postal vouchers.