딥러닝에 기반한 다양한 환경에서의 색상 및 문자 인식

Abstract

Since the previous decade, Deep Learning revolutionized the field of Artificial Intelligence by providing it with an immense performance boost. The performance continues to get better with the increase in amount of data. Deep Learning has been made possible due to the advances in hardware technologies that has enabled fast computation of large number of complex hidden layers in the neural network. The exemplary performance of Deep Learning has been the motivation for its implementation in this thesis. In this thesis, Deep Learning is applied for Character Recognition in tire images. Previously, the localization of individual characters was done using horizontal and vertical histogram projections of the image. In an attempt to increase the accuracy of localization, this thesis proposes to implement Deep Learning algorithms, particularly Convolutional Neural Network (CNN). This thesis also classifies color by averaging all pixels in the R, G, and B channel of a color image. Then these three calculated average values are fed into a feed forward neural network for classification. However, this is not enough due to the fact that the color of an object changes with varying illumination conditions. The human color perception has the ability to perceive a relatively constant color despite changing illumination conditions. This phenomenon of identifying the color even with varying illumination is known as Color Constancy. Unfortunately, machines do not recognize this concept of color constancy which leads to an inaccurate color classification. For example, a banana is yellow when the illumination is white but its color changes with the change in the color of illumination. This thesis automatically estimates the color of illumination using CNN.