机械类文献翻译对车牌定位的研究

机械专业中英文文献翻译

英文原文

1.Introduction

Nowadays license plate recognition becomes akey technique to many automated transport systemssuch as road traffic monitoring, automaticpayment of tolls on highways or bridges and parkinglots access control. License plate location is anessential and important stage in this technique,and it has received considerable attention.

Researchers have found many diverse methods of license plate location. Rodolfo and Stefano (2000) devised a method based on vector quantization (VQ). VQ image representation is a quadtree representation by the specific coding mechanism,and it can give a system some hints about the contents of image regions, and such information boosts location performance. Park et al. (1999)used neural networks to locate license plate. Neural networks can be used as filters for analyzing small windows of an image and deciding whether each window contains a license plate, and their inputs are HSI values; a post-processor combinesthese filtered images and locates the bounding boxes of license plates in the image. Besides neural networks, other filters have been considered too. For example, some authors used line sensitive filters to extract the plate areas. License plates are identified as image areas with high density of rather thin dark lines or curves. Therefore, localization is handled looking for rectangular regions in the image containing maxima of response to these line filters, which is computed by a cumulative function (Luis et al., 1999). Plate characters can be direct identified by scanning through the input image and looking for portions of the image that were not linked to other parts of the image.If a number of characters are found to be in a straight line, they may make up a license plate (Lim et al., 1998). Fuzzy logic has been applied to the problem of locating license plate by Zimic et al. (1997). The authors made some intuitive rules to describe the license plate, and gave some membership functions for the fuzzy sets ??bright?? and ??dark??, ??bright and dark sequence?? to getthe horizontal and vertical plate positions. But this method is sensitive to the license plate color and brightness and needs much processing time. Using color features to locate license plate has been studied by Zhu et al. (2002) and Wei et al. (2001), but these methods are not robust enough to the different environments. Edge features of the car image are very important, and edge density can be used to successfully detect a number plate location due to the characteristics of the number plate. Ming et al. (1996) developed a method to improve the edge image by eliminating the highest and lowest portions of the edge density to simplify the whole image. But some of the plate region identity will be lost in this method.

This paper further researches the subject of license plate location. The rectangle license plate contains rich edge and texture information, so we consider it in its edge image but very different to Ming et al. (1996). We first enhance the original car image to boost up the plate area, then extract the vertical edge image using Sobel operator, and then remove the background curves and noise in the edge image, and finally slide a rectangle window to search the plate in the residual image and

机械专业中英文文献翻译

segment it out from the original car image. Section2 describes our method of license plate location, and it contains four parts: image enhancement, vertical edge extraction, background curve and noise removing, plate search and segmentation. Experiments with three sets of car images are performed in Section 3. Section 4 gives the discussion and conclusions.

1. The proposed method for license plate location

All the input car images have 384 · 288 pixels and 256 gray levels, and an example image is given in Fig. 1. The license plate of the car consists of several characters (such as Latin letters, Arabic numerals, etc.), so the plate area contains rich edge information. But sometimes the background of the car image holds much edge information too. There are two facts that attract our attention: one is that the background areas around the license plate mainly include some horizontal edges; the other is that the edges in the background are mainly long curves and random noises, whereas the edges in the plate area cluster together and produce intense texture feature. If only the vertical edges are extracted from the car image (although the plate will lose a little horizontal edge information, this little loss is to be valuable) and most of the background edges are removed, the plate area will be isolatedout distinctly in the whole edge image. Thus we propose to locate the license plate in its vertical edge image as the following four stages.

2.1. Image enhancement

In Fig. 1, the gradients in the license plate area are much lower than those in the contour areas of the car, which is caused by the car shadow in the dazzling sunshine. The car images captured in the gloomy days or dim nights often bring out weak gradients in plate areas too. A few vertical edges will appear in the plate areas, if we extract edge images directly from these car images. Therefore it is important to enhance the car images firstly.

The local areas that need to be enhanced in a car image have low variances. Here we use Ii,j to denote the luminance of the pixel Pi,j (row:0 6 i < 288, column: 0 6 j < 384) in the car image, and use I1i;j to denote the luminance in the enhanced image. We let Ii,j and I1 i;j satisfy Eq. (1), where Wi,j is a window centered on pixel Pi,j, IW i;j and rW i;j are the mean luminance and standard deviation of the pixels in the window Wi,j, I0 andr0 are the expected mean and standard deviation,respectively.

I1i,j??0(Ii,j?Iwi,j)?I0?wi,j （1）

In order to represent the local information better, the size of the window should be smaller than the estimated size of the plate. In this paper, we select a 48 · 36 rectangle as the window Wi,j and thus 8 · 8 windows can cover over the whole 384 · 288 car image. Let I0 be equal to IW i;j and r0 be a constant independent of pixel Pi,j. Now we need to know the values IW i;j and rW i;j at each pixel. Computing out all the values is not advisable, and we can use the bilinear interpolation algorithm to get them. First we cut the car image into 8 · 8 blocks equably; and then compute out the IW i;j

机械专业中英文文献翻译

and rW i;j values at the vertexes of blocks, where i = 36m, j = 48n,m,n = 0,1,2, . . . ,8; finally compute out every IW i;j and rW i;j by the bilinear interpolation Eqs. (2) and (3) (Fig. 2), where 36m 6 i < 36(m + 1), 48n 6 j < 48(n + 1), cx =(j ?? 48n)/48, and cy = (i-36m)/36.

Iwi,j?(1?cy)[(1?cx)IwA?cxIwB]?cy[(1?cx)IwD?cxIwD]（2）

?w?(1?cy)[(1?cx)?w?cx?w]?cy[(1?cx)?w?cx?w] （3）

i,jABCDIf σW i;j is close to zero (such as only a dark or bright area), Eq. (1) will give out a large value. But we should not enhance such a local area. If σW i;j is high enough (for example σW i;j ≥60), the enhancement is unnecessary too. So the enhancement Eq. (1) is improved into Eq. (4) inpractice.

Ii?,j?f(?wi,j)?(Ii,j?Iwi,j)?Iwi,j （4）

where f（r ） is an enhancement coefficient(shown in Fig. 3) defined by Eq. (5). Most σW i;j s of

W

i;j

the plate areas which need enhanced are around 20. So we let the function f be equal to 1 when rW i;j =0 or rW i;j ≥60, and be equal to 3 (as 20 · 3 = 60) when rW i;j =20.

3?if0??Wi,j?20?22(?Wi,j?20)?1??400?3f(?Wi,j)??if20??Wi,j?60 （5）

2?(?Wi,j?20)2?1?1600?1if??60Wi,j??The enhanced car image is shown in Fig. 4. And we can see that the license plate region has been strengthened. If the plate is well illuminated and the image is in balance, the process will not change the contrast of the plate (f（σwi,j=1,if σwi,j≥60）.

2.2. Vertical edge extraction

We select the vertical Sobel operator (in Fig. 5) to detect the vertical edges, because the simple operator costs us a little computational time. Convolve the car image with this Sobel operator to get the vertical gradient image. Compute the mean of the absolute gradient values in the image and multiplyit by a coefficient as a threshold (for example 4Grad), or compute the gradient histogram and find a gradient at a certain percentage (for example 75%) of the gradient distribution as a threshold.Use this threshold and apply nonmaximum suppression in horizontal direction in the gradient image, and we get the vertical Sobel edge image shown in Fig. 6.

2.3. Background curve and noise removing

From Fig. 6, we can see that there are many long background curves and short random noise