Proceedingsof ICSP2000 Car License Plates Detection from Complex Scene Da-shan Gao Jie Zhou Department of Automation, Tsinghua University Beijing 100084, P.R.China [email protected]zhoLi~ie(cc,infa.aii.tsinlrhua.edu.cii Abstract: In this paper, we present a novel approach to extract car license plate from complex image without reading attempt. After an algorithm of segmentation, a series of candidate regions are obtained first. Then a confidence value based on the geometrical features of license plates is given to each candidate region and merge operation under some rules is taken. E xperim ental results show that the algorithm is robust in dealing with different conditions such as poor illumination and distortion of image generated by different visual angle. 1. Introduction Automati c recognition of car license plates plays an important role in traffic surveillance systems. Such systems, which are applied in parking areas, highways, bridges and tunnels, can help a human operator and improve the overall quality of a service. Any situation requiring the automatic control of the presence and license number may represent a potential application. Recently, we have seen quite a few computer-vision- based systems that recognize the license plates [l-91. Most existing systems focus on the development of a reliable optical character recognizer (OCR). Howe ver prior to the recognition an OCR system performs, the license plate has to be extracted from a variable of scenes. Since there are problems such as poor ambient lighting problem, visual angle, image distortion and so on, sometimes the car license plate is difficult to be extracted. Many techniques have been reported in previous researches. Hough Transform for line detection was proposed in [3] on the assumption that the shape of license plate is defined by lines. Combining extraction of license plates with character recognition by BP neural networks was used in [4]. 5,6] used neural networks (NN) with some features in car license plate such as color and so on. Vector Quantization methodology and distributed genetic algorithm was used in [ 7 ] an d [8], respectively. Although the algorithm proposed in [9] is robust for recognition of inclined license plates due to different visual angle, it depends on the high quality acquired by a special CCD and a set of strict prior know ledge. In this paper we proposed a novel method to extract car license plates from a complex scene by considering both the distributive regulation of the characters in a license plate and the geometrical features of a license plate. In our approach, we first present a segmenting algorithm, looking for the candidate regions that probably contain characters in a proper size. Then we give each candidate region a confidence value to measure its likelihood to be a license plate and combine these regions according to some rules to get a higher confidence value. Then the car license pl ate can be found to have highest value. 2. Car license plate detecti on 2.1. Outline of the algorithm In this algorithm we present a technique for the location and extraction of car license plates in complex scenes. As schematized in Figure 1 the algorithm works in four major steps: preprocessing, extraction of candidate regions, morphological processing, endowing confidence value and region merge. 2.2. Preprocessing As the image that contains car license plates are acquired in a real environment under uncontrolled 0-7803-5747-7/00/$10.00Q2000IEEE.
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
8/4/2019 Car License Plates Detection From Complex Scene
contrast. In order to reduce the undesired effects and
enhance the contrast, histogram equalization for
contrast enhancem ent can be used when necessary.Histogram Equalization is a technique in image
processing, through which we desire to take a given
input image into an output image with equally manypixels at every gray level (a flat histogram) Ell]. In
practice, we set a threshold observed in a set ofimages and equalize the image whose accumulativehistogram is below this threshold. After histogramequalization a contrast enhancement is applied by a
sigmoid transform function,
f ( x ) = 1+exp(--a * ( x - d ) )
where c, d and a are constant which determine the
maximum value, center and shap e of the function.
C
(1)
2.3. Extraction of candidate regions
Considering the characters in a car license plate
always have a distinctive gray level to the background
of the license plate, which is to say that a car license
plate have a relative high contrast, we can get thefollowing features of a license plate in its gradientimage.
Firstly, the average gradient value of the region that
contains a license plate is high because of the intensevariations in it, as is mentioned in [I]. The size of thewindow where the local average gradient value is
calculated can be set to correspond to the size oflicense plates in the majority of images acquiredthrough a CCD camera.
Secondly, the variance of the gradient image of a
license plate region is relative low because there are anumber of edges of characters in it. So the variance is
calculated in a window of certain size to distinguish a
license plate region from a long edge with only highcontrast.
Dividing the average gradient value by the local
variance of gradient image at each pixel, we perform abinary method successively and get some candidateregions, which probably contain a car license.
2.4. Morphological Processing
After the image are segmented by thresholding,
there may be. some noise in the image such as isolated
dots, long vertical or horizontal stripes and so on . Soan opening operation of morphological processing[113, in which a Dilation operation is performed after
an erosion operation, is applied in order to reduce the
undesired effect of noise, smooth the edges of the
candidate regions and to separate the regions which
should be separated.
2.5.Geom etrical Criteria an d Confidence Value
To detect a license plate from a complex scene is a
kind of simplified detection of a text line in an imagein a sense [lo]. It is easier because of the fixed
geometrical structure of license plates, which appearsin almost the same shape of rectangle and contains
characters with the same number. So we specify some
geometrical criteria and confidence functions, thevalue of which is from 0 to 1, based on the internalfeatures of a license plate to depict the likelihood
between a candidate region and a license plate region.In the following, we discuss these internal features
respectively.Area. The area o f a region is defined as the number
of its pixels. As a recognizable license plate, it mustcontain quite a few pixels. So the larger the area of aregion is, the higher the confidence v alue will be.
Elongation. A license plate can be regard as ahorizontal rectangle with particular ratio of width andheight. Even though sometimes it is distorted in aimage from different visual angle, it still can be
bounded by a skew rectangle with approximate rationof width and height. With this prior knowledge, we
find the two axis of a region through K-L transformand make a minimum rectangle to bind the region.The elongation feature is defined as the ratio of widthand height of the rectangle. The more approximate to
the ratio of a real license plate region the elongation is,
the higher a confident value will be given.Density. It is defined as the ratio between the
region area and the area of a bounding rectanglediscussed above. In general, a license plate region is
fully filled. So the index permit to detect sparsely
filled regions, which is given a low confidenc e value.
Proximity to the image frame (PIF). Theproximity index is defined as the distance between the
pixels of the region and the image frame, normalized
with respect to the corresponding image size. In manycases of application in traffic control system, a car is
the focus in a image or almost is. So a license platecan .be found in relative center region of an image.
This feature is introduced to identify such noisy
regions that often appear along the border of theimage.
8/4/2019 Car License Plates Detection From Complex Scene