AUGMENTATION PROGRAM OF PUPILOGRAPHY DATA FOR CONSTRUCTING NEURAL NETWORK CONVOLUTIONAL MODELS FOR DETERMINING THE FUNCTIONAL STATE OF HUMAN FATIGUE

The article is devoted to the development and use of a program for augmenting pupillograms and constructing neural network convolutional models designed to determine the functional state of human fatigue. A promising area of research is the analysis of the dynamics of the human pupillary response to a light stimulus to assess the level of mental workload, stress and fatigue of a person. However, the use of machine learning methods for processing pupillographic data is often limited due to the insufficient amount and variability of labeled data, which reduces the accuracy and generalizing ability of the models. Generating synthetic data taking into account the physiological patterns of pupillary responses will improve the quality of model training and improve their ability to determine the functional state of human fatigue. To solve this problem, the work used an experimental dataset collected by a team of KNITU-KAI researchers using a special hardware and software complex. The set includes 384 pupillogram images measuring 640x480 pixels of 2 classes corresponding to human fatigue and wakefulness. The following methods were selected for dataset augmentation: jitter, drift, time distortion, and value averaging. Python was selected as the programming language for developing the program. The developed program includes the following functionality: loading txt files, selecting a directory for uploading files, selecting data augmentation methods, visualizing pupillograms, augmenting data, verifying generated data, uploading generated data in txt and png formats. Using the program allowed us to expand the dataset from 384 images to 831. To evaluate the augmentation efficiency and the quality of the resulting dataset, two convolutional networks based on the ResNet model with the same hyperparameter values were trained. The first network was trained on the original dataset, and the second on the augmented one. The values of the Accuracy, Precision, Recall, and F1-measure metrics were calculated. The results of the study showed the efficiency of using pupillogram data augmentation to build convolutional models. The obtained results can be used as a basis for further research in the field of determining the functional state of human fatigue.