CN108830857A - A kind of adaptive Chinese character rubbings image binaryzation partitioning algorithm - Google Patents

Abstract

The invention discloses an adaptive binarization segmentation algorithm for images of inscriptions with Chinese characters. The method is divided into four steps: (1) preprocessing by median filtering; (2) extracting red components; (3) morphological operations with Find the best background estimate; (4) Otsu segment the binary image. The invention belongs to the technical field of image segmentation of Chinese character inscriptions, which preserves the stroke features of Chinese characters and enhances the details of characters. Aiming at degraded images of historical inscriptions, the invention proposes an image binarization adaptive segmentation algorithm based on background estimation and non-uniform illumination degradation , the novelty of the proposed method lies in finding an optimal background estimate based on blind/no-reference image spatial quality assessment, and experimental results show that the method is capable of more accurate character segmentation for degenerated Chinese characters.

Description

An Adaptive Binarization Segmentation Algorithm for Images of Inscriptions of Chinese Characters

technical field

The invention belongs to the technical field of image segmentation of Chinese-character inscriptions, and specifically refers to an adaptive binary image segmentation algorithm for Chinese-character inscriptions.

Background technique

An important way for China to preserve its own history and culture is the memory written on stones - inscriptions. At the same time, rubbings are an important part of ancient Chinese books, and they are the main source for people to learn and research history. The digitization of rubbing documents is a new way to promote and inherit Chinese traditional art, and a new way of thinking to protect stone cultural relics. On the other hand, rubbings of extant ancient books often lose their visual quality. With the passage of time, due to the influence of many factors such as its storage conditions, humidity, and pollution, after the rubbings are copied into digital images, it is difficult to separate the foreground and background due to reasons such as different contrast, high noise, and enhanced background intensity. In addition to low contrast, background noise often introduced textured stone, and the aging of paper materials, it is a challenging task to extract clean Chinese characters from the original image of ancient rubbings. But the above is also a key step for any further automatic document image analysis, such as layout analysis, character recognition, etc. The brightness distribution in the process of rubbing image acquisition is uneven, which will affect the quality of the image.

In order to improve the quality of character segmentation, the inventors have studied a large number of methods and technologies, among which the most important preprocessing step is text binarization, which converts the document image from a grayscale or color image to a binary image. The foreground and background of ancient document images are separated on the basis that white pixels represent the foreground and the foreground is represented by black pixels. One of the simplest and most efficient image processing techniques that can be used to separate the foreground and background layers of a document image is thresholding, a number of thresholding techniques that can be classified as global and local thresholding algorithms, multi-thresholding methods, and adaptive thresholding techniques. Global thresholding is preferred when images have the same contrast on the background and foreground, they are evenly illuminated, and the target and background differ greatly. Locally adaptive thresholding is used to recover foreground pixels in document images. In general, choosing an algorithm for degraded inscription images is a very difficult process. Due to the existence of complex degradation, many experimental results show that the traditional method of uniform illumination processing of weak target images has disadvantages such as incomplete separation of target and background or loss of target, and low processing efficiency. Therefore, the inventors proposed a new adaptive algorithm to process rubbing images, that is, using blind/no-reference image spatial quality assessment, using the Natural Image Statistics (NSS) model framework to locally normalize the brightness coefficient, and using the "non- nature". To correctly segment low-contrast color images, a morphological Top-Hat operator with disk-shaped structural elements and adaptive pixels is employed to exploit the red component of the difference between rubbing structures. To reduce noise, median filtering is applied to the shading corrected image.

In general, degraded document images are due to background noise and variations in contrast and brightness. Document images with degraded shadows are more common because camera documents are more susceptible to lighting changes. There are many existing algorithms that try to segment the foreground and background when scanning a text document, but thresholding is of one form or another and standard tools for another region, such as Bernsen's Adaptive Thresholding is estimated based on the field of each pixel. Use local maxima and minima to construct a local contrast image. Then, a sliding window is applied on this image to determine a local threshold. The prior art proposes a phase-based binarization model for segmenting ancient document images, and develops a real generation tool called PhaseGT to simplify and speed up the process of generating real ancient document images. Recently, an active contour evolution algorithm has also been proposed in the prior art, which is used to automatically generate our active contour model based on the intrinsic geometric measures of the document image, image contrast, i.e., by local maxima and minima of the image, initializing the image ; Finally, an average threshold can also be generated and finally binarized. As observed by the inventors, most binarization methods are based on intuitive observation of gray levels between characters and background, regardless of adaptive selection thresholds for degraded document images. To overcome these difficulties, this scheme proposes an adaptive method that applies different methods to segment characters from degraded document images.

Contents of the invention

In order to solve the above existing problems and preserve the stroke features of Chinese characters while enhancing the character details, the present invention proposes an image binarization adaptive segmentation algorithm based on background estimation and non-uniform illumination degradation for degraded historical rubbing images. The novelty of the method is to find an optimal background estimate based on blind/no-reference image spatial quality assessment, which is divided into four steps: (1) preprocessing with median filtering; (2) extraction of red components; ( 3) Morphological operation to find the best background estimation; (4) Otsu segment binary image. Experimental results show that this method can perform more accurate character segmentation on degenerate Chinese characters.

The technical scheme that the present invention adopts is as follows: a kind of self-adaptive Chinese character stele post image binarization segmentation algorithm, uses color image to develop a kind of robust algorithm, is used for segmenting Chinese rubbing image from background, comprises the following steps:

1) Reprocessing using median filtering, which allows a lot of high spatial frequency detail to pass through while being very effective at removing noise on images that are less than half the pixels in smooth neighborhoods;

2) Extract the red component;

3) Morphological image processing operations so that if a minimum BRISQUE is found, the optimal diameter Thr* of the disc can be found;

4) Use Otsu's to segment the binary image.

Further, the morphology described in step 3) is mathematical morphology, and the basic idea of using mathematical morphology for image processing is to use structural elements with a certain shape (basic elements with a certain structural shape, such as rectangles, circles or rhombuses) etc.) to detect the target image, by analyzing the target area of the image and the effectiveness of the structural elements in the filling method, the relevant information about the shape and structure of the image is obtained, and they are used to achieve the purpose of image analysis and recognition.

Furthermore, the structural elements are a key point in morphological image processing. Different structural elements determine the analysis and processing of various geometric information in the image, and also determine the amount of calculation in the data conversion process. Therefore, the structural elements Analysis is an important part of image edge detection; the size and shape of structural elements will affect image edge detection; small-sized structural elements have weak denoising ability, but they can detect precise edge details; large-sized structural elements have stronger The denoising ability, but the detected edge is rougher; more importantly, the structural elements of different shapes have different processing capabilities for the edges of different images; among them, the grayscale image can be regarded as a set of two-dimensional points, dilation and the corrosion operation can be expressed as follows:

The Top-hat algorithm can be divided into Top-hat algorithm and Bot-hat algorithm according to the different components of the opening operation and the closing operation. The Top-hat algorithm is applied to the image and expressed as TH:

Apply the Top-hat algorithm to the image and denote it as BH:

BH(x,y)=(f⊙S-f)(x,y) (4)

In the equation, f(x, y) is the original grayscale image, S(x, y) is the structural element, the Top-hat transformation extracts the foreground information by the difference between the original image and its opening operation, and the Bot-hat transformation Background information is suppressed by the difference between the original image and its closing operation.

Further, the BRISQUE described in step 3) is a general-purpose no-reference image quality assessment algorithm based on spatial image statistical features, which is based on the following theoretical premise: natural images have certain regularity, and the visual features of human eyes follow the regularity Evolution, In the prior art, Anish Mittal and other researchers found that the normalized luminance coefficients of natural images in the spatial domain have statistical properties and follow a unit Gaussian distribution. This function is affected by image distortion, and different distortions have different effects on the distribution. Based on the above research results, we propose a BRISQUE no-reference image quality assessment algorithm based on spatial domain statistical features. For a given grayscale image of size M*N, the brightness normalization coefficient of each pixel satisfies the following:

In the formula: i=1,2,...,M; j=1,2,...,N; c is a constant, c=1; K=L=3; i M=...1,2,,; j N= …1,2,,; c is a constant, c=1; K=L=3; μ(i,j) and σ(i,j) are mean and standard deviation; ω={ω} _kj ∣k = -K, -K+1, ..., K, L = -L, -L+1, ..., L is the sampling and normalization of the two-dimensional Gaussian equation; the BRISQUE algorithm uses the brightness normalization coefficient as a quality-related feature to evaluate Image Quality. In contrast to other non-reference quality assessments, the use of image features eliminates the need for various complex transformations. Therefore, the algorithm has the advantages of simple calculation and time saving under the premise of similar accuracy. On the other hand, the decorrelation process of image brightness ignores the effect of brightness on the quality of the test image.

Further, step 4) uses Otsu's segmentation binary image to include a binarization segmentation algorithm, including Jaccard coefficient, false positive rate FPR and false negative rate FNR three parameters in the binary segmentation algorithm for hierarchical measurement, false positive The rate FPR shows the degree of under-segmentation, and the false-negative rate FNR shows the degree of over-segmentation; the Jaccard coefficient measures the similarity between finite sample sets and is defined as the size of the intersection divided by the size of the union of the sample sets. For binary images, calculate The intersection of binary images A and B is divided by the union of A and B, and the Jaccard coefficient can be calculated using the following formula:

The false positive rate FPR and false negative rate FNR are defined as follows:

where FP is the number of false positives, white in real images and black in binarized images, FN is the number of false negatives, TN is the number of true negatives, and N=FP+TN is the total number of negatives.

By adopting the above scheme, the present invention achieves the following beneficial effects: the present invention retains the stroke features of Chinese characters while enhancing character details, and proposes an image binarization adaptive segmentation algorithm based on background estimation and non-uniform illumination degradation for degraded historical rubbing images , the novelty of the proposed method lies in finding an optimal background estimate based on blind/no-reference image spatial quality assessment. The method is divided into four steps: (1) Preprocessing with median filtering; (2) Extracting red components; (3) morphological operations to find the best background estimate; (4) Otsu segmentation of binary images. Experimental results show that this method can perform more accurate character segmentation on degenerate Chinese characters.

Description of drawings

Fig. 1 is a flow chart of the method of the present invention;

Fig. 2 is the rubbing image of conventional Chinese tablet image;

Fig. 3 is the histogram of Fig. 2;

Fig. 4 is the objective function value diagram of Fig. 2;

Figure 5 is a comparison chart of image segmentation results.

Detailed ways

The following will clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only some of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

Embodiment, a kind of self-adaptive binarization segmentation algorithm of Chinese character inscription post image in this embodiment, uses color image to develop a kind of robust algorithm, is used for segmenting Chinese rubbing image from background, comprises the following steps:

1) Reprocessing using median filtering, which allows a lot of high spatial frequency detail to pass through while being very effective at removing noise on images that are less than half the pixels in smooth neighborhoods;

2) Extract the red component;

3) Morphological image processing operations so that if a minimum BRISQUE is found, the optimal diameter Thr* of the disc can be found;

4) Use Otsu's to segment the binary image.

A flowchart of the proposed segmentation method is shown in Figure 1.

Experimental analysis and results,

To evaluate and test the proposed method, this scheme uses a database of Chinese inscription images from the University of California, Berkeley East Asian Library ( http://ucblibrary4.berkeley.edu:8088/xtf3/search?rmode=stonerubbings&identifier=&title=&name =&text=&date=&startdate=15&subject=&height=&width=&material=&script=&enc_provenance=). This program uses the Jaccard similarity coefficient, FPR (false positive rate) and FNR (false negative rate) to compare with the classic Otsu algorithm to evaluate the performance and quality of the system of this program. As shown in Figure 2-4, the rubbing image of the conventional Chinese inscription image is dark, the histogram shows that the image contrast is very low, and the image quality of historical documents is degraded. Before estimating the foreground, it is necessary to apply a median filter to remove noise. A median filter considers each pixel in the image in turn and looks at its sphere to determine if it is representative of its sphere. The neighborhood is chosen as a 3×3 pixel square. This is a very small neighborhood compared to the image size, and the final image size is 371×1260 pixels. Afterwards, a morphological opening is performed on the estimated foreground using the function imopen in Matlab.

The binarization method introduced above is applied to our test set, which consists of images of old Chinese rubbing post documents with several degradations and structural complexities. We present in this example the results of applying previous methods to our collection of images. For objective evaluation, we use the Jaccard coefficient, false positive rate (FPR) and false negative rate (FNR) for stratified measurement. FPR shows the degree of under-segmentation, and FNR shows the degree of over-segmentation. The Jaccard coefficient measures the similarity between finite sample sets and is defined as the size of the intersection divided by the size of the union of the sample sets. For binary images, it calculates the intersection of binary images A and B divided by the union of A and B . The Jaccard coefficient can be calculated using the following formula:

The false positive rate FPR and false negative rate FNR are defined as follows:

where FP is the number of false positives, white in real images and black in binarized images, FN is the number of false negatives, TN is the number of true negatives, and N=FP+TN is the total number of negatives.

The performance of the segmentation algorithm is shown in Table 1.

Table 1 Quantitative measurements of segmentation

(d) in Figure 5 is the Chinese character segmentation result obtained by using the adaptive segmentation algorithm of this scheme, (c) in Figure 5 is the result obtained by the Ostu' segmentation algorithm obtained, and (d) in Figure 5 is the basic real binary image , refers to the actual binary image with all noise and degradation factors manually removed. In Table 1, it can be noticed that the best results for low-contrast document images are obtained by the proposed method. In Table 1, the Jaccard coefficient is significantly higher than the Ostu method. OTSU's global thresholding method misclassifies some text pixels, while misclassifying dark background pixels as text pixels. The experimental results show that the background removal algorithm proposed in this scheme can achieve more accurate restoration of various Chinese rubbing images than the Ostu method. For low-contrast images of Chinese inscriptions, it performs well.

However, since the threshold is applied globally, the threshold for some weak handwriting may cause the handwriting to be broken.

In summary, the images of Chinese inscriptions obtained from rubbings have the characteristics of many blurred details and poor effects, so more details may be lost in the traditional processing process. Preprocessing is an important stage in image processing, especially in the case of Chinese ancient book image segmentation applications. An efficient image preprocessing algorithm will improve the accuracy of segmentation algorithms and reduce misclassification. The invention proposes an adaptive segmentation algorithm aiming at the binarization of degraded rubbing image. Subjective and objective evaluation methods are used to judge the efficiency of our algorithm. Experimental results show that estimating the image background by morphological operations is an adaptive choice to find the optimal diameter of the disk. However, since our background estimation algorithm does not consider the illuminance relationship in different scenes, it is possible to introduce slight flicker for other applications when the scene changes significantly. In the future, the method will be tested in an OCR (Optical Character Recognition) application to test the readability of the proposed method in degraded documents.

In addition, this invention has been strongly supported by the National Natural Science Foundation of China (61472173), the Natural Science Foundation of Jiangxi Province (20161BAB202042), and the Jiangxi Provincial Education Commission Funding Project (GJJ151134).

The above descriptions are only examples of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the content of the description of the present invention, or directly or indirectly used in other related technical fields, shall be The same reasoning is included in the patent protection scope of the present invention.

Claims (5)

1. A self-adaptive Chinese character inscription post image binarization segmentation algorithm is characterized in that, a kind of robust algorithm is developed using color image, is used for segmenting Chinese rubbing image from background, comprises the steps: 1) Reprocessing using median filtering, which allows a lot of high spatial frequency detail to pass through while being very effective at removing noise on images that are less than half the pixels in smooth neighborhoods; 2) Extract the red component; 3) Morphological image processing operations so that if a minimum BRISQUE is found, the optimal diameter Thr* of the disc can be found; 4) Use Otsu's to segment the binary image. 2. a kind of self-adaptive Chinese character inscription post image binarization segmentation algorithm according to claim 1, it is characterized in that, step 3) described morphology is mathematical morphology, utilizes mathematical morphology to carry out the basic idea of image processing It uses structural elements with a certain shape to detect the target image, and analyzes the effectiveness of the structural elements in the target area of the image and the filling method to obtain information about the shape and structure of the image, and use them to achieve the purpose of image analysis and recognition. 3. a kind of self-adaptive Chinese character rubbing post image binarization segmentation algorithm according to claim 2, it is characterized in that, described structural element is a key point of morphology image processing, and different structural elements have determined each in the image. The analysis and processing of various geometric information also determines the amount of calculation in the data conversion process, so the analysis of structural elements is an important content of image edge detection; the size and shape of structural elements will affect image edge detection; small-sized structural elements have weak denoising ability, but they can detect precise edge details; large-scale structural elements have stronger denoising ability, but the detected edges are rougher; more importantly, different shapes of structural elements have different effects on different images. The edges of have different processing capabilities; among them, the grayscale image can be regarded as a set of two-dimensional points, and the dilation and erosion operations can be expressed as follows: (f⊙S)(x,y)=min{g(x-k,y-l)|(k,l)∈S} (2) The Top-hat algorithm can be divided into Top-hat algorithm and Bot-hat algorithm according to the different components of the opening operation and the closing operation. The Top-hat algorithm is applied to the image and expressed as TH: Apply the Top-hat algorithm to the image and denote it as BH: BH(x,y)=(f⊙S-f)(x,y) (4) In the equation, f(x, y) is the original grayscale image, S(x, y) is the structural element, the Top-hat transformation extracts the foreground information by the difference between the original image and its opening operation, and the Bot-hat transformation Background information is suppressed by the difference between the original image and its closing operation. 4. a kind of self-adaptive Chinese character inscription post image binarization segmentation algorithm according to claim 1, it is characterized in that, step 3) described BRISQUE is a kind of general no reference image quality assessment algorithm based on spatial image statistics , for a given grayscale image of size M*N, the brightness normalization coefficient of each pixel satisfies the following: In the formula: i=1,2,…,M; j=1,2,…,N; c is a constant, c=1; K=L=3; i M=…1,2,…,; j N =...1,2,...,; c is a constant, c=1; K=L=3; μ(i,j) and σ(i,j) are mean and standard deviation; ω={ω} _kj ∣k=-K, -K+1, ..., K, L = -L, -L+1, ..., L is the sampling and normalization of the two-dimensional Gaussian equation; the BRISQUE algorithm uses the brightness normalization coefficient as a quality-related feature to evaluate image quality. 5. a kind of self-adaptive Chinese character inscription post image binarization segmentation algorithm according to claim 1, it is characterized in that, step 4) comprises binarization segmentation algorithm in the segmentation binary image of using Otsu, binarization segmentation algorithm The three parameters including Jaccard coefficient, false positive rate FPR and false negative rate FNR are used for stratified measurement. The false positive rate FPR shows the degree of under-segmentation, and the false negative rate FNR shows the degree of over-segmentation; the Jaccard coefficient measures the similarity between limited sample sets. , and is defined as the size of the intersection divided by the size of the union of the sample set. For binary images, calculate the intersection of binary images A and B divided by the union of A and B. The Jaccard coefficient can be calculated using the following formula: The false positive rate FPR and false negative rate FNR are defined as follows: where FP is the number of false positives, white in real images and black in binarized images, FN is the number of false negatives, TN is the number of true negatives, and N=FP+TN is the total number of negatives.