Use this URL to cite or link to this record in EThOS:
Title: Saliency-based search free car license plate localisation
Author: Safaei, Amin
ISNI:       0000 0004 6423 2548
Awarding Body: University of Surrey
Current Institution: University of Surrey
Date of Award: 2017
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Restricted access.
Access from Institution:
Advances in intelligent transportation systems (ITS) has great societal impact. Automatic license plate recognition (ALPR) or automatic number plate recognition (ANPR) systems have brought valuable insight for the red-light or speed limit enforcement, electronic payment systems, and traffic surveillance. These systems comprise of four separate phases: image acquisition, localisation, segmentation, and character recognition. Blocked or covered licence plates (LPs) and images taken under bad environmental situations, lightings, size and orientation variations are some of the bottlenecks of ALPR systems. In real-time operations e.g. ITS, ALPR process has to operate rapidly and accurately. Real-time applications of ALPR systems require the fast and accurate detection of LPs becomes even one single miss detection or lack of detection might cause problems. That is why the localisation phase of ALPR systems seem to be the most crucial one, determining the speed and the accuracy of the whole system. Popular search-dependent algorithms cannot afford to keep up with the challenges of ALPR systems, due to their computational complexity. This motivates development of search-free methods in this thesis. The proposed methods can be categorised into five groups; algorithms based on: saliency map and local variance; incorporating hierarchical saliency; multiple LPL (license plate localisation) via likelihood estimation; incorporating the differences in distributions, such as by negentropy, and finally through a deep convolution neural network approach for LPL. The first proposed algorithm does not need to search the image and it is based on estimations of saliency and local variance (using L1 − norm). It also uses Gabor function to exploit directionality for validating and choosing the right LP. In this algorithm, traditional saliency map detection (SMD) method locates the cars in images in a very fast, search-free approach. Next proposed approach modifies the traditional SMD method by means of directionality and a new definition of saliency, called saliency based ALPR (SB-ALPR). In addition, L1 −norm is used to better identify the pattern of characters and numbers in the LP regions. Experimental results show an average 94.77% accuracy, 61.52 ms execution time and 40.2 ms processing time per LP, which is a remarkable achievement in comparison to the performance of the previous algorithm using traditional SMD method. A search-free LP localisation algorithm on the basis of 3-D Bayesian saliency estimation is the next proposed method. In this algorithm, 3-D objects are traced by means of object/shadow detection and removal and Bayesian method for object recognition. This algorithm consists of three fundamental phases: object/shadow detection and object recognition. For the object detection phase, the background image and the moving objects are detected. To eliminate the shadows, a new approach, in which we discriminate the shadows from their corresponding objects, is exploited and for tracking purpose, the relations between the objects in sequential frames are determined via Bayesian method. The results show that, unlike the previous algorithm, in this proposed method the image backgrounds are more accurately subtracted and the shadows eliminated by an accuracy of approximately 70%. Object recognition phase is also subjoined to the algorithm, which performed well. In another novel approach the distance between the distributions by means of negentropy is combined with the saliency based ALPR. We then utilise the Bayesian decision making method for selecting the LP from the detected candidates. This proposed algorithm shows an accuracy of 96% and a computation time of 80ms per plate, which is an outstanding improvement over the classic touchstone techniques. In the last proposed method a deep convolution neural network is successfully used for LPL. Following this approach the network learns the necessary filters at each layer and generates new patterns by convolving the filters with the images in previous layers.
Supervisor: Cheong Took, Clive Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available