Utilising computer vision to solve arithmetics on Android
Year: 2018
Type: dissertation
Abstract: The objective of this thesis was to create a native Android application, which uses machine learning to solve hand-written addition equations from images captured with a device’s camera. Additionally, the application was required to use a machine learning model trained with a pre-existing dataset and optimally be able to complete the given task on high end devices without noticeable latency. \n \nA pretrained mobile first deep convolutional neural network from the MobileNets family was chosen as the machine learning model. The MobileNet model had relatively low requirements for processing power, making it possible to run inference on a mobile device in real time. The neural network was trained with an augmented MNIST handwritten digits dataset combined with a custom dataset for the addition operator class. \n \nImage preprocessing and character extraction functionality was implemented on the client to allow the machine learning model to classify hand-written characters with a satisfactory accuracy. Most of the preprocessing steps which have been applied to the MNIST dataset are also performed in real time for all camera input by the application. \n \nThe outcome of the thesis is an Android application which can be used to scan and solve hand-written addition equations from a device’s camera feed in real time. Testing on handwritten characters proved that the application could be used with reasonable accuracy.
Source: Theseus (Ammattikorkeakoulujen)
Author: Matti Mäki-Kihniä
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2796765862
