RECOGNITION OF ENGLISH CAPITAL LETTERS BASED ON CONVOLUTIONAL NEURAL NETWORK

The article solves the problem of developing a convolutional neural network model for recognizing handwritten letters of the English alphabet. The main approaches to character recognition are considered. It is noted that traditional methods of optical character recognition based on rules and statistical models are giving way to more flexible and efficient neural network approaches. To implement this approach, the following steps were performed: obtaining the initial data for analysis, building the architecture of the neural network model, training the model and evaluating the results of its operation. To build the model, a publicly available EMNIST dataset on the Kaggle platform was selected, containing 88,800 images of handwritten letters of the English alphabet. For clarity, an example of images from the dataset is provided. The architecture of the neural network model is described. The MATLAB modeling environment was used as a platform for its training. The Adam optimizer with an initial learning rate of 0.001 and a maximum number of epochs of 10 was used to train the neural network. The batch size was set at 64, which ensured a balance between the learning rate and the quality of weight updates. The model was trained using the trainNetwork function, which accepted a prepared set of images and labels. Each training epoch included an accuracy calculation, which made it possible to monitor the progress of model construction. The accuracy of the neural network model after training was 92%. To evaluate the performance of the neural network image classifier, an error matrix was constructed, the analysis of which allowed us to conclude that the neural network model has difficulties recognizing the letters "i" and "l" since these letters are very similar and, based on different handwritings and small noises, it becomes impossible to determine without additional information which of the letters is presented in the image. To test the model's robustness to external distortions, a study was conducted that included adding random Gaussian noise to the images from the test set. After adding noise, the accuracy of the neural network classification decreased from 92% to 88.27%. This indicates that the model maintains relatively high accuracy despite the presence of noisy data. The constructed model demonstrated high accuracy of recognition of images of English handwritten letters. Despite the decrease in accuracy when adding noise and distortions, the model demonstrated good generalization ability. This indicates its adequacy and the possibility of effective practical use.