CN108830857B - Self-adaptive Chinese character copy label image binarization segmentation method - Google Patents

Abstract

The invention discloses a self-adaptive Chinese character inscription image binarization segmentation method, which comprises the following four steps: (1) preprocessing by median filtering; (2) extracting red components; (3) morphological operations to find the best background estimate; (4) otsu segments the binary image. The invention belongs to the technical field of Chinese character poster image segmentation, retains stroke characteristics of Chinese characters and enhances character details, and provides an image binarization self-adaptive segmentation algorithm based on non-uniform illumination degradation of background estimation aiming at degraded historical poster images.

Description

Self-adaptive Chinese character copy label image binarization segmentation method

Technical Field

The invention belongs to the technical field of Chinese character copy label image segmentation, and particularly relates to a self-adaptive Chinese character copy label image binarization segmentation method.

Background

An important way for China to retain its own historical culture is to write a memory-inscription on a stone. Meanwhile, the rubbing is an important component of ancient Chinese books and is a main source of learning and research history of people. The digitalization of rubbing literature is a new way for developing and propagating the traditional Chinese art and is a new idea for protecting the stone cultural relics. On the other hand, existing ancient book rubbings tend to lose visual quality. Over time, due to the influence of many factors such as storage conditions, moisture, pollution, etc., after the rubbing is copied into a digital image, the foreground and the background are difficult to be separated due to different contrast, loud noise, enhanced background intensity, etc. Besides low contrast and background noise, which is often introduced into the texture stone, there is also aging of the paper material, so extracting clean chinese characters from the original image of the tombstone is a challenging task. But the above is also a key step in performing any further automatic document image analysis, such as layout analysis, character recognition, etc. The brightness distribution in the rubbing image acquisition process is uneven, and the quality of the image is influenced.

In order to improve the quality of character segmentation, the inventors studied a number of methods and techniques, of which the most important preprocessing step is text binarization, converting a document image from a grayscale or color image to a binary image, and separating the foreground and background of an ancient document image on the basis that background information is represented by white pixels and the foreground by black pixels. One of the simplest and most efficient image processing techniques can be used to separate the foreground and background layers of a document image, namely thresholding, many thresholding techniques that can be categorized as global and local thresholding algorithms, multi-thresholding methods and adaptive thresholding techniques. Global thresholds are preferred when the image has the same contrast on the background and foreground, their illumination is uniform, and the object and background are very different. The locally adaptive threshold is used to recover foreground pixels in the document image. In general, selecting an algorithm for degraded post images is a very difficult process. Due to the existence of complex degradation, a plurality of experimental results show that the traditional weak target image illumination uniform processing method has the defects of incomplete target and background separation or target loss, low processing efficiency and the like. The inventors therefore propose a new adaptive algorithm to process topological images, i.e. local normalization of luminance coefficients using a natural image statistics (NSS) model framework using blind/no-reference image spatial quality estimation, using "non-natural" quantified by model parameters. For correct segmentation of low contrast color images, the morphological Top-Hat operator is used with disc-shaped structuring elements and adaptive pixels, taking advantage of the red component of the difference between the topographies. To reduce noise, median filtering is applied to the shading-corrected image.

In general, a degraded document image is generated due to background noise and variations in contrast and brightness. Shadow-degraded document images are more common because camera documents are more susceptible to lighting changes. There are many algorithms that attempt to segment the foreground and background when scanning text documents, but the threshold is some form and standard tool for another region, such as Bernsen's adaptive threshold, which is estimated from the domain of each pixel. Local maxima and minima are used to construct a local contrast image. A sliding window is then applied over the image to determine the local threshold. The prior art provides a binary model for segmenting ancient document images based on phases, and a real generation tool, called PhaseGT, is developed to simplify and accelerate the generation process of the real ancient document images. Recently, an active contour evolution algorithm is also proposed in the prior art, which is used for automatically generating an active contour model of the user and initializing an image according to the intrinsic geometric measurement and the image contrast of a document image, namely the local maximum value and the local minimum value of the image; finally, an average threshold may also be generated and finally binarized. As observed by the inventors, most binarization methods are based on an intuitive observation of the gray level between the character and the background, regardless of the adaptive selection threshold of the degraded document image. To overcome these difficulties, the present solution proposes an adaptive method that applies different methods to segment characters from degraded document images.

Disclosure of Invention

In order to solve the existing problems, the stroke characteristics of Chinese characters are kept, the character details are enhanced, aiming at degraded historical tombstone images, the invention provides an image binarization self-adaptive segmentation algorithm based on non-uniform illumination degradation of background estimation, the provided method is novel in that an optimal background estimation based on blind/non-reference image space quality evaluation is found, and the method comprises the following four steps: (1) preprocessing by median filtering; (2) extracting red components; (3) morphological operations to find the best background estimate; (4) otsu segments the binary image. Experimental results show that the method can perform more accurate character segmentation on the degraded Chinese characters.

The technical scheme adopted by the invention is as follows: a self-adaptive Chinese character stele image binarization segmentation method develops a robust algorithm by using a color image and is used for segmenting a Chinese rubbing image from a background, and comprises the following steps:

1) using a median filtering process that allows a large amount of high spatial frequency detail to pass while very effectively eliminating noise on images less than half a pixel in a smooth neighborhood;

2) extracting red components;

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

4) binary images were segmented using Otsu.

Further, the morphology in step 3) is mathematical morphology, and the basic idea of image processing using mathematical morphology is to use structural elements with certain shapes (basic elements with certain structural shapes, such as rectangles, circles, diamonds, etc.) to detect a target image, obtain information about image morphology and structure by analyzing effectiveness of structural elements in an image target area and a filling method, and use them to achieve the purpose of image analysis and recognition.

Further, the structural element is a key point of morphological image processing, and different structural elements determine analysis and processing of various geometric information in an image and also determine the calculation amount in the data conversion process, so that the analysis of the structural element is an important content of image edge detection; the size and the structural shape of the structural elements can influence the image edge detection; small-sized structural elements have weak denoising capability, but they can detect precise edge details; the large-size structural elements have stronger denoising capability, but the detected edges are rougher; more importantly, the structural elements with different shapes have different processing capacities on the edges of different images; where the grayscale image can be viewed as a set of two-dimensional points, the dilation and erosion operations can be represented as follows:

(f⊙S)(x，y)＝min{g(x-k，y-l)|(k，l)∈S} (2)

the Top-hat algorithm can be divided into a Top-hat algorithm and a Bot-hat algorithm according to different components of the open operation and the closed operation, and is applied to an image and expressed as TH:

the Top-hat algorithm is applied to the image and is denoted BH:

BH(x，y)＝(f⊙S-f)(x，y) (4)

in the equation, f (x, y) is an original grayscale image, S (x, y) is a structural element, Top-hat transform extracts foreground information by the difference between the original image and its opening operation, and Bot-hat transform suppresses background information by the difference between the original image and its closing operation.

Further, the BRISQUE in step 3) is a general reference-free image quality evaluation algorithm based on spatial image statistical characteristics, and the algorithm is based on the following theoretical premise: in the prior art, Anish Mittal and other researchers find that the normalized luminance coefficient of the natural image in the spatial domain has statistical characteristics and conforms to unit Gaussian distribution. This function is affected by image distortion, with different distortion pairs being assigned different effects. Based on the research results, a BRISQUE non-reference image quality evaluation algorithm based on spatial domain statistical characteristics is provided. For a given grayscale image of size M × N, the luminance normalization factor for each pixel satisfies the following:

in the formula: 1,2, …, M; j ═ 1,2, …, N; c is a constant, c is 1; k ═ L ═ 3; i M ═ … 1,2,; j N ═ … 1,2,; μ (i, j) and σ (i, j) are mean and standard deviation; ω ═ ω_klI K-K, -K +1, …, K, L-L, -L +1, …, L is a sample of a two-dimensional Gaussian equationAnd a normalized parameter; the BRISQUE algorithm uses the luminance normalization coefficient as a quality-related feature to evaluate the image quality. The use of image features eliminates the need for various complex transformations compared to other non-reference quality assessments. Therefore, the algorithm has the advantages of simple calculation and time saving on the premise of approximate precision. On the other hand, the decorrelation process of the image brightness ignores the effect of brightness on the test image quality.

Further, in the step 4), Otsu segmentation binary images comprise a binarization segmentation algorithm, the binarization segmentation algorithm comprises three parameters of Jaccard coefficient, false positive rate FPR and false negative rate FNR for layered measurement, the false positive rate FPR shows an under-segmentation degree, and the false negative rate FNR shows an over-segmentation degree; the Jaccard coefficient measures the similarity between a finite sample set and is defined as the size of the intersection divided by the size of the union of the sample sets, for a binary image, the intersection of binary images A and B is calculated divided by the union of A and B, and the Jaccard coefficient can be calculated using the formula:

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

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

By adopting the scheme, the invention has the following beneficial effects: the invention keeps the stroke characteristics of Chinese characters and enhances the character details, and provides an image binarization self-adaptive segmentation algorithm based on non-uniform illumination degradation of background estimation aiming at degraded historical tombstone images, wherein the novel point of the provided method is to find an optimal background estimation based on blind/non-reference image space quality evaluation, and the method comprises the following four steps: (1) preprocessing by median filtering; (2) extracting red components; (3) morphological operations to find the best background estimate; (4) otsu segments the binary image. Experimental results show that the method can perform more accurate character segmentation on the degraded Chinese characters.

Drawings

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

FIG. 2 is a rubbing image of a conventional Chinese tombstone image;

FIG. 3 is the histogram of FIG. 2;

FIG. 4 is a graph of the objective function values of FIG. 2;

fig. 5 is a map of the image segmentation result.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

The embodiment provides a self-adaptive Chinese character tombstone image binarization segmentation method, which develops a robust algorithm by using a color image and is used for segmenting a Chinese rubbing image from a background, and comprises the following steps:

1) using a median filtering process that allows a large amount of high spatial frequency detail to pass while very effectively eliminating noise on images less than half a pixel in a smooth neighborhood;

2) extracting red components;

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

4) binary images were segmented using Otsu.

A flow chart of the proposed segmentation method is shown in fig. 1.

The analysis and the results of the experiments are as follows,

to evaluate and test the proposed method, the present solution uses a database of chinese signature images from the east asian library of the berkeley division, university of california (r) ((r))http://ucblibrary4.berkeley.edu:8088/xtf3/ searchrmode＝stonerubbings&identifier＝&title＝&name＝&text＝&date＝&startdate＝15&subje ct＝&height＝&width＝&material＝&script＝&enc _ provenance ═ g). The scheme compares Jaccard similarity coefficient, FPR (false positive rate) and FNR (false negative rate) with a classic Otsu algorithm to evaluate the performance and quality performance of the system of the scheme. As shown in fig. 2-4, the rubbing image of the conventional chinese tombstone image is dark, the contrast of the histogram display image is low, and the image quality of the historic document is reduced. Before estimating the foreground, it is necessary to apply a median filter to remove noise. The median filter considers each pixel in the image in turn and looks at its domain to determine whether it represents its domain. The neighborhood is chosen to be a square of 3 x 3 pixels. This is a very small neighborhood compared to the image size, with a final image size of 371 × 1260 pixels. Then, using the function imopen in Matlab, a morphological opening operation is performed on the estimated foreground.

The binarization method described above is applied to our test set consisting of old chinese post document images with several degradations and structural complexities. We present in this embodiment the results of applying the previous method to our favorite images. For objective evaluation, we performed hierarchical measurements using three parameters, Jaccard coefficient, False Positive Rate (FPR) and False Negative Rate (FNR), with FPR showing the degree of under-segmentation and FNR showing the degree of over-segmentation. The Jaccard coefficient measures the similarity between a finite sample set and is defined as the size of the intersection divided by the size of the union of the sample sets, for a binary image it computes the intersection of the binary images A and B divided by the union of A and B. The Jaccard coefficient can be calculated using the following formula:

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

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

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

TABLE 1 quantitative measurement results of the segmentation

	Jaccard	FPR	FNR
				Algorithm of scheme	0.6009	0.0348	0.3781
OTSU algorithm	0.3948	0.7433	0.3188

Fig. 5 (d) is a chinese character segmentation result obtained using the present scheme adaptive segmentation algorithm, (c) in fig. 5 is a result obtained by the obtained Ostu' segmentation algorithm, and (d) in fig. 5 is a substantially true binary image, which refers to an actual binary image from which all noise and degradation factors are manually removed. In table 1, it can be noted that the best results for low contrast document images are obtained by the present approach. In Table 1, the Jaccard coefficient is significantly higher than the Ostu method. The global thresholding method of OTSU misclassifies some text pixels while misclassifying dark background pixels as text pixels. The experimental result shows that the background elimination algorithm provided by the scheme can realize more accurate restoration on various Chinese poster images than the Ostu method. It performs well for low contrast chinese tombstone images.

However, since the threshold is applied globally, for some weak handwriting, the handwriting may be corrupted.

In conclusion, the Chinese tombstone images obtained by rubbing have the characteristics of much fuzzy details, poor effect and the like, so that more details can 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 pre-processing algorithm will improve the accuracy of the segmentation algorithm and reduce misclassification. The invention provides a binarization self-adaptive segmentation algorithm for a degraded poster 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 take into account the luminance relationships in different scenes, it is possible to introduce slight flicker for other applications in case of a significant change in the scene. In the future, the method will be tested in OCR (optical character recognition) applications to test the readability of the proposed method in degraded documents.

In addition, the invention obtains the great support of China national science fund (61472173), Jiangxi province science fund (20161BAB202042) and Jiangxi province educational committee subsidy project (GJJ 151134).

The above description is only an embodiment of the present invention, and not intended to limit the scope of the present invention, and all modifications of equivalent structures and equivalent processes, which are made by the present specification, or directly or indirectly applied to other related technical fields, are included in the scope of the present invention.

Claims (4)

1. A self-adaptive Chinese character inscription image binaryzation segmentation method is characterized in that a robust algorithm is developed by using a color image and is used for segmenting a Chinese rubbing image from a background, and the method comprises the following steps:

1) using a median filtering process that allows a large amount of high spatial frequency detail to pass while very effectively eliminating noise on images less than half a pixel in a smooth neighborhood;

2) extracting red components;

3) morphological image processing operations so that if a minimum BRISQUE is found, the optimal diameter Thr of the disk can be found; the morphology is mathematical morphology, the basic idea of image processing by using the mathematical morphology is to detect a target image by using structural elements with certain shapes, obtain related information of the image morphology and structure by analyzing the effectiveness of structural elements in an image target area and a filling method, and realize the purposes of image analysis and identification by using the related information;

4) binary images were segmented using Otsu.

2. The adaptive Chinese character tokenization image binarization segmentation method as claimed in claim 1, wherein the structural element is a key point of morphological image processing, different structural elements determine analysis and processing of various geometric information in an image and determine a calculated amount in a data conversion process, so that analysis of the structural element is an important content of image edge detection; the size and the structural shape of the structural elements can influence the image edge detection; small-sized structural elements have weak denoising capability, but they can detect precise edge details; the large-size structural elements have stronger denoising capability, but the detected edges are rougher; more importantly, the structural elements with different shapes have different processing capacities on the edges of different images; where the grayscale image can be viewed as a set of two-dimensional points, the dilation and erosion operations can be represented as follows:

(f⊙S)(x，y)＝min{g(x-k，y-l)|(k，l)∈S} (2)

the Top-hat algorithm can be divided into a Top-hat algorithm and a Bot-hat algorithm according to different components of the open operation and the closed operation, and is applied to an image and expressed as TH:

the Top-hat algorithm is applied to the image and is denoted BH:

BH(x，y)＝(f⊙S-f)(x，y) (4)

in the equation, f (x, y) is an original grayscale image, S (x, y) is a structural element, Top-hat transform extracts foreground information by the difference between the original image and its opening operation, and Bot-hat transform suppresses background information by the difference between the original image and its closing operation.

3. The adaptive binarization segmentation method for Chinese character post images as claimed in claim 1, wherein the BRISQUE in step 3) is a general reference-free image quality evaluation algorithm based on spatial image statistical characteristics, and for a given grayscale image with size M x N, the brightness normalization coefficient of each pixel satisfies the following condition:

in the formula: 1,2, …, M; j ═ 1,2, …, N; c is a constant, c is 1; k ═ L ═ 3; i M ═ … 1,2, …; j N ═ … 1,2, …; μ (i, j) and σ (i, j) are mean and standard deviation; ω ═ ω_kl-K +1, …, K, L-L, -L +1, …, L } are parameters of the sampling and normalization of a two-dimensional gaussian equation; the BRISQUE algorithm uses the luminance normalization coefficient as a quality-related feature to evaluate the image quality.

4. The adaptive Chinese character inscription image binarization segmentation method as claimed in claim 1, wherein in the step 4) segmentation of the binary image by Otsu comprises a binarization segmentation algorithm, the binarization segmentation algorithm comprises three parameters of Jaccard coefficient, false positive rate FPR and false negative rate FNR for hierarchical measurement, the false positive rate FPR shows an under-segmentation degree, and the false negative rate FNR shows an over-segmentation degree; the Jaccard coefficient measures the similarity between a finite sample set and is defined as the size of the intersection divided by the size of the union of the sample sets, for a binary image, the intersection of binary images A and B is calculated divided by the union of A and B, and the Jaccard coefficient can be calculated using the formula:

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

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

Images