FASHION MNIST IMAGE RECOGNITION BY DEEP LEARNING METHODS

Abstract

A review of modern methods for recognizing image objects has shown that deep learning algorithms are successfully used and provide high quality. An example is the quality of character recognition of the MNIST set, which is close to 100%. For another dataset, which is also popular in the implementation of deep learning algorithms, namely, the Fashion MNIST set of clothing items and accessories, such a high recognition quality has not yet been achieved. The paper presents the results of data recognition Fashion MNISТ. Models of a feedforward neural network and convolutional network are considered. The software implementation of deep learning algorithms is carried out, namely, a multilayer feedforward network (FNN) and a convolutional neural network (CNN) are considered. The Python language, the TensorFlow and Keras libraries are used. The Keras library allows you to simplify TensorFlow function calls. A typical workflow in Python Keras is as follows: loading the necessary modules, loading data, preprocessing, breaking them into training, test and validation parts; creating a model with an indication of the architecture, and more. A numerical experiment was carried out to recognize clothing items by means of FNN. Automatic tuning of network hyperparameters has been performed. The recognition quality on test data is 0.89. The hyperparameter tuning did not significantly improve the quality. The low quality of recognition is also explained by the use of a very simple neural network model. Improved results have been achieved using convolutional neural networks. The best recognition results obtained in the work is 91.26%, but the known best recognition quality result is 94%. It is advisable to continue the work on improving the results of recognition of images of Fashion MNIST data, and the software that has been developed can be used to recognize other data. With a creative approach to image object recognition in Python using the Keras Tensorflow libraries and others, it is a very promising direction for practical application.