Browsing by Author "Kalken M."

Now showing 1 - 2 of 2
  • Thumbnail Image
    ItemOpen Access
    Automatic Handwritten Digit Recognition on Images Using Machine Learning Methods
    (2022) Kalken M.
    With the transformation to digital information exchange, many papers, including invoices, taxes, notes and surveys, historical data, and test results, still require handwriting. In this context, handwritten digit recognition is required, which is a computer-automated method for decoding records. To achieve this purpose, machine learning mostly employs neural networks, with the convolution neural network being the most common. The primary objective of this dissertation is to develop an automatic technique for detecting handwritten digits by determining the optimal hyperparameter of a convolutional neural network and analyzing the impact of various hyperparameters on network performance. To do this, several values of each hyperparameter were chosen within a specified range using the MNIST database for training and testing. As a result, a large amount of data was collected in which a very high influence of hyperparameter selection on the performance of the neural network was determined, and a 99.2% accurate pure CNN design was achieved.
  • Thumbnail Image
    ItemOpen Access
    HANDWRITTEN OPTICAL CHARACTER RECOGNITION: IMPLEMENTATION FOR KAZAKH LANGUAGE
    (СДУ хабаршысы - 2021, 2021) Kalken M.
    Abstract. Many documents, including as invoices, taxes, memoranda, and surveys, historical data, and test replies, still require handwriting with the transformation to digital information interchange. Handwritten text recognition (HTR), which is an automatic approach to decode records using a computer, is required in this aspect. For this proposal, I present a study of the implementation of optical recognition algorithms for handwritten text in the Kazakh language, using a recently collected database. The database, called the Kazakh Autonomous Handwritten Text Dataset (KOHTD), contains more than 140,335 segmented images of handwritten exam papers. As an algorithm, I used the proposed model by Harald Scheidl, which consists of several layers of neural networks and an CTC decoder. The trained model by putting an interval of Ir = 0.01 and a batch size of 60 showed effective results with indicators of about 85% accuracy.