Scholarly record

ON APPLICATION OF HEAT EQUATION IN DEEP LEANING RECOGNITION OF GOOGLE STREET VIEW NUMBERS

Vasyl Martsenyuk

Abstract

Recognition of the numbers is a necessary condition for the development of 3D street viewers. Using information about the position of the viewer relative to the detected site by self-driving car will make it possible to adjust the path and increase the accuracy of its implementation. The unmanned device (unmanned aerial vehicle or self-driving car) is equipped with a video camera or camcorder system. In the process of detecting house numbers, the image from the camera contains noises and redundant information. The preceding recognition step involves preliminary processing of the image obtained from the optical system. One of the techniques for removing noise from an incoming image is to perform an anti-aliasing operation. In this work we consider the filter presented by Peron and Malik, which is a modification of the heat equation. The choice of the method is due to the presence of a boundary detector, which allows us to save significant edges on the image when filtering noise. This is necessary, for example, for algorithms of recognition based on contour analysis. The recognition of filtered image is implemented in R with help of keras package. As a training set we use the Street View House Numbers (SVHN) Dataset, which is a real-world image dataset with minimal requirement on data preprocessing and formatting. It was shown that the primary filtering of test images allows us to increase accuracy of number recognition.

Publication Impact Profile

• Captures
• Mendeley - Readers: 4

PLUMX see details

DOI resolver: doi.org/10.5593/sgem2019/2.1/s08.128

Publication details

Back to publication list