Improvement of Character Segmentation for Indonesian License Plate Recognition Algorithm using CNN

Abstract

Recognition of vehicle license plate based on computer vision is very useful for replacing human eye from manually identifying license plate. In practice, the algorithm of licence plate recognition needs to be robust to various orientations, noises and illuminations of captured plates. Conventionally, one of the challenging processes is segmenting the characters of detected plate. The segmented characters are extracted to perform recognition. Thus, performance of character segmentation affects the final result. This research aims to perform character segmentation of Indonesian license plate by applying detection of character using Convolutional Neural Network (CNN) and sliding window with bounding box refinement. In this proposed method, CNN is used to distinguish character and non-character region. To feed regions to CNN, sliding window technique is applied. The final bounding boxes are finally refined to increase the accuracy. And the developed model was tested on 130 Images of Indonesian vehicle license plate which contain 982 characters in total, and yielded 87.06% of accuracy.