Improved license plate recognition for degraded vehicle imags in CCTV surveillance systems

Abstract

Intelligent transportation systems (ITSs) aim at improving transportation safety and productivity. To that end, these systems make use of advanced technologies such as automatic license plate recognition (LPR), which has recently attracted considerable amount of research attention. Although significant progress has been made during the last few decades, it is still difficult to achieve effective LPR for real-world imagery, mainly due to the challenging nature of the capturing conditions of vehicle images. Indeed, it is typical for vehicle images to have low spatial resolution and to contain a substantial amount of (motion) blur and noise, due to the frequent use of closed circuit television (CCTV) camera in long distance. In order to handle the aforementioned problems, we pay attention to the following observations in developing a practical LPR system than can be applied to real-world environment. Firstly, we have prior knowledge about target license plates such as color configurations of license plates and location of characters. Secondly, recent CCTV captures color image, which can provide rich information for object recognition. Thirdly, since content of CCTV is not a still image, we can make use of more than one frame of same vehicle for LPR. In light of these observations, we propose an improved LPR system for recognizing degraded vehicle images in real-world applications. Our contributions in this thesis are threefold. First, we propose an improved character segmentation method using ROI calibration. Even if not all the characters are found at once in initial character detection due to the severe degradation, we calibrate ROI based on prior knowledge and find the missing characters in estimated region. Second, we propose an enhanced character feature which exploits Gabor wavelet with color information, which is proven to be robust to low-resolution object recognition. Third, we incorporate image sequence of same vehicle to find vehicle identification num...