Real-time fashion items classification using TensorflowJS and ZalandoMNIST dataset
Nguyen, Kha (2020)
Nguyen, Kha
2020
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-202004235681
https://urn.fi/URN:NBN:fi:amk-202004235681
Tiivistelmä
The last decade has marked a significant growth in Deep Learning, driven mainly by the rapid development of Graphics Processing Unit (GPU) which provide sufficient computing power for deep neural network training. However, AI platforms are usually centrally hosted on the computing instances provided by Cloud Service Providers and therefore leads to undesirable networked latency.
Acknowledging the problem, Google released TensorFlowJS, an open-sourced library for Machine Learning in an attempt to provide an interactive in-browser Machine Learning environment for researchers and practitioners to perform various machine learning tasks.
This thesis will uncover the fundamentals of deep learning by training deep artificial neural network models using FashionMNIST dataset and use TensorFlowJS to build a real-time hand-drawing image classification based on trained models. Traditionally, MNIST is the most popular dataset for validating an algorithm in deep learning; however, in April 2017, the pioneer of Artificial Intelligence Ian Goodfellow stated that MNIST is being overused by the Machine Learning community and that researchers and practitioners should move on to harder datasets. These facts inspired Zalando's research team to replace the obsolete MNIST with FashionMNIST – a dataset consists of 70,000 examples of Zalando's article images in 28x28 grayscale format.
To demonstrate the usability and effectiveness of in-browser Machine Learning platform, the author constructed a neural network model using Keras and train it with FashionMNIST. The parameters of the neural network were saved in text files and loaded to the browser using TensorFlowJS. The thesis project also made use of Canvas, an HTML5 feature which allows users to draw fashion items in the browser and convert the depicted image to TensorFlow input for prediction.
The outcome of the project is a functional web app which features a hand drawing area and a pie chart representing the prediction made by the neural network model.
Acknowledging the problem, Google released TensorFlowJS, an open-sourced library for Machine Learning in an attempt to provide an interactive in-browser Machine Learning environment for researchers and practitioners to perform various machine learning tasks.
This thesis will uncover the fundamentals of deep learning by training deep artificial neural network models using FashionMNIST dataset and use TensorFlowJS to build a real-time hand-drawing image classification based on trained models. Traditionally, MNIST is the most popular dataset for validating an algorithm in deep learning; however, in April 2017, the pioneer of Artificial Intelligence Ian Goodfellow stated that MNIST is being overused by the Machine Learning community and that researchers and practitioners should move on to harder datasets. These facts inspired Zalando's research team to replace the obsolete MNIST with FashionMNIST – a dataset consists of 70,000 examples of Zalando's article images in 28x28 grayscale format.
To demonstrate the usability and effectiveness of in-browser Machine Learning platform, the author constructed a neural network model using Keras and train it with FashionMNIST. The parameters of the neural network were saved in text files and loaded to the browser using TensorFlowJS. The thesis project also made use of Canvas, an HTML5 feature which allows users to draw fashion items in the browser and convert the depicted image to TensorFlow input for prediction.
The outcome of the project is a functional web app which features a hand drawing area and a pie chart representing the prediction made by the neural network model.