1) Convert Original Image to YUV Color Space Image

RGB color space is complex in describing the color pattern and has redundant information between each component. Since pixel values in RGB color space are highly correlated, RGB color space is converted into other color spaces. The YUV color model defines a color space in terms of one luminance and two chrominance components. Because of the low-contrast images we choose have similar color between character and background, meanwhile, the luminance information can get a better result than colorful information for further processing, so, we convert the input image (see Figure 1) into YUV color space (luminance + chrominance), and use only luminance (Y) channel (see Figure 2) for further processing.

2) Enhance the Above Y Channel Image to Increase Image Contrast

Owing to the text images we choose are low-contrast, so, in order to get a better result, we first select some methods to enhance the image, comparing with multi-scale retina algorithm, adjust image intensity, histogram equalization and contrast-limited adaptive histogram equalization, finally finds that contrast-li- mited adaptive histogram equalization (see Figure 3) can get a good results than other algorithms.

3) Use Sobel Edge Detector Operator to Detect the Above-Enhanced Image

After doing some enhancement for the image, we need to select a kind of edge operator to detect the image edge. In the experiment, we can see Sobel operator [7] [8] (see Figure 4) is able to detect all the text edge and meanwhile remove the non-text edge.

4) Use Morphology Method (MM) and Eight Neighborhood Processing Method (ENPM) to Enhance Edge Information

Observing the Sobel detector results, we found the edge density is very weak, so we decide to adopt morphology method [9] and eight neighborhood processing method proposed by us to enhance edge information. About method ENPM, it is that the value in edge map gets by its eight neighborhood value, only if there is one appearance in its neighborhood, its value will be replaced by one. Figure 4 displays Sobel operator detector results, and Figure 5 and Figure 6 respectively show the results processed by morphology method and eight neighborhood processing method. Obviously, edge density is enhanced after using morphology method and eight neighborhood processing method.

Following is the algorithm about ENPM (see Table 1).

CCA could be regarded as a graph algorithm, where subsets of connected components are uniquely labeled based on heuristics about feature consensus, i.e., color similarity and spatial layout. In the implementation of CCA, syntactic pattern recognition methods are often used to analyze the spatial and feature consensus, and to define text regions. Considering the complexity of fine turning the syntactic rules, a new trend is to perform CCA with statistical models.

The basic steps of the connected-component text extraction algorithm are given below. 1) Convert the input image to YUV color space (luminance + chrominance), use only luminance channel for further processing; 2) Convert the image gray scale using only Y channel; 3) Compute the edge image for Y channel gray image; 4) Sharpen the edge image by convolving it with sharpening filter; 5) Compute horizontal and vertical projection files considering the sharpened edge image as the input intensity image; 6) Segment the candidate text regions based on adaptive threshold values calculated for vertical and horizontal projections; 7) Perform gap filling to eliminate possible non-text regions [10] .

Through the above steps, we can use CCA method to remove part overlap blocks from connect component.

In Figure 7(a), a lot of non-text region or overlap regions are been selected, which is useless and time-consuming for us, so we need to take some measures to remove these overlap blocks. Figure 7(b) shows results after removing part

overlap blocks. After removing the overlap regions, we need to extract the remaining regions except for the regions gained by CCA.

The type of split overlap region is as follows. In which the yellow color region the first block and the white color region is the second block. According to dual law, there are four types (see Figure 8) for overlap which the second block is on the border or inside in the first block.

Figure 9 presents the extracted results from image processed by CCA removed part overlap blocks.

1) First Layer Judge (FLJ)

Reading 200 text images, and using the above method to gain a lot of connected component regions, by observing the regions difference between text regions and non-text regions, we find that regions whose row or column is less than 22, or the value for row divide column is more than 10 or less than 0.1 generally would be non-text regions. The results see in Figure 10 and Figure 11. In Figure 9, there are 36 blocks extracted from the images, while after using first layer judge method, the number of blocks reduced to 14 in Figure 10.

2) Second Layer Judge (SLJ)

a) Get the Edge Detector Image

Edge is the distinct characteristic which can be used to find possible text areas. The text is mainly composed of the strokes in horizontal, vertical, up-right, up- left direction, so it can be considered that the region with higher edge strength in these directions is the text region. In order to get a better edge strength, we made a lot of experiments to see which style of computing would be better.

In Figure 12 the number a from g shows the block extracted from the Y channel image and a1 from the g1 shows the correspondence binary image.

Figure 13 gives the edge detection results with four angles. We find that 45 with 135 averages are better than others. For saving time, undoubtedly, we choose these two directional edge maps to represent the density and strength for edge enhancement.

b) Based Edge Map Image to Detection Unsupervised Text Regions

To some extent, the text has weak and irregular texture property, so we can look text as a special texture. We employ the statistical features in the edge maps to capture the texture property. The features are mean, standard deviation, energy, inertia, local homogeneity and correlation of edge maps [11] .

After feature computing, we get 5 features form the feature vector representation for each block, then use Support Vector Machine Classifier Algorithm to classify the feature vectors into two clusters: text regions and background regions (see Table 2).

Figure 14 is the flow chart of an overall process for a specific image to realize automatically text position.

Text region merges mainly rely on the location of the text region, finding the neighborhood region, and compute the distance in near neighborhood regions. Then merge with those regions which their distance is small. Figure 15 shows the start state before merging to end state which had successfully located all the text regions.

Figure 17 shows the experimental results of various kinds of images selected from www image or advertisements. In Figure 17(a), Figure 17(b) and Figure 17(d), Chinese texts with different size in images are successfully detected. Figure 17(c) demonstrates that this approach is effective for the texts in a gloomy environment.

Here we collect a new dataset with 200 images of various low-contrast images. In addition, in order to evaluate our method, we compute the precision rate and recall rate, meanwhile, makes comparisons with other methods.

The performance of each algorithm has been evaluated based on its precision rate, an average run time obtained. The precision and recall rates are calculated as [12]

PrecisionRate = Correctlydetected Correctlydetected + Falsepositives × 100 % (1)

RecallRate = Correctlydetected Correctlydetected + Falsenegatives × 100 % (2)

We compute the results in Table 3. After counting the recognition results, we get the Table 4.