Method for detecting and recognizing characters in power pole signboard images

Abstract

The present application relates to the image processing technical field in machine vision, especially relates to the high-voltage transmission line inspection image processing field in power system, and specifically relates to a kind of character detection and identification method in power pole information board image.The power pole information board image is preprocessed first;Then the feature points of the power pole information board image after preprocessing are obtained, and the character candidate region is obtained;Remove the non-character region in the character candidate region, obtain single character region;Single character region is merged into text line;Text line is recognized by character.This application makes full use of the advantages of MSER detection natural scene text, gives full play to the advantages of Imbinarize-MSER local feature multi-scale detection, candidate region stability is good, affine invariance, and the character detection and identification of the power pole information board collected, facilitate monitoring and management the running state of transmission line, provide security for the safety of transmission line.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image processing in machine vision, in particular to the field of high-voltage transmission line inspection image processing in power system, and specifically to a method for character detection and recognition in power pole information board image. BACKGROUND

[0002] The power transmission line is a crucial component of the power grid system, which plays a decisive role in the reliable, long-term, safe and stable operation of the entire power grid as the backbone of the power grid system, and the long-term effective operation of the power grid system is directly related to the healthy and stable development of the national economy. The information in the power pole (tower) number plate is the "identity card" of the power line, which not only contains the voltage level, location, number and other information of the line, but also plays a role in promoting power knowledge and safety warning. The power pole information board greatly facilitates the relevant work of power inspection and maintenance, especially the daily inspection of unmanned aerial vehicles and power robots. If a line fault occurs, the fault location and position can be quickly and accurately found through the power pole information board, improving work efficiency. With the increase in the number of power grids, the number of power poles and towers also increases year by year, and the information in the power pole number plate is rich, which plays a very important role in ensuring the safety of the power grid.

[0003] For character detection and recognition of power pole information board in natural scenes, in addition to the influence of strong electric field, illumination, blur, non-text and other factors, it is also affected by the arrangement direction of adjacent texts in the image.

[0004] Currently, the power pole information board image character detection and recognition method relies on various Chinese character and number templates, which not only requires the establishment of a large number of template libraries, but also greatly increases the time and computational complexity of the traversal process of each template. The recognition system based on neural network has the problems of large model size, large sample quantity, complex training, and low real-time performance.

[0005] The character detection algorithm and OCR recognition technology based on Imbinarize-MSER and mathematical morphology can well describe the consistency of the internal color of the text and overcome the interference caused by noise and affine transformation to a certain extent. The maximum stable extreme region is a subset of the extreme region. Since the maximum stable extreme region is closed for continuous coordinate transformation and monotonic transformation of image brightness, it has good robustness to translation transformation, angle transformation, scale transformation, illumination transformation, and rotation transformation. Mathematical morphology can simplify image data, maintain their basic shape characteristics, and remove irrelevant structures. OCR technology is relatively mature and has high recognition rate. SUMMARY

[0006] In view of the problems of rich text information and complex background of power pole information board images, the application provides a character detection and recognition method for power pole information board images, which uses an MSER improved algorithm, geometric features and mathematical morphology for detection, an undirected graph for merging regions and an OCR recognition technology, performs simulation experiments through a Matlab platform, detects and locates by using a local feature detection algorithm, and has high recognition accuracy.

[0007] The application is implemented by the following technical scheme: a character detection and recognition method for power pole information board images, comprising the following steps:

[0008] S1: performing pretreatment on collected power pole information board images by using an affine transformation and a cropping algorithm;

[0009] S2: extracting feature points of the pretreated power pole information board images by using an Imbinarize-MSER algorithm, and obtaining a character candidate region;

[0010] S3: removing non-character regions in the character candidate region by using geometric features and mathematical morphology, and obtaining a single character region;

[0011] S4: merging the single character region into a text line by using an undirected graph;

[0012] S5: performing text recognition on the text line by using an OCR technology.

[0013] As a further improvement of the technical scheme of the application, in step S1, the affine transformation specifies a 3x3 numerical matrix, each power pole information board image is regarded as a pixel point arranged in rows and columns, [x y 1]=[x0 y0 1]xT, P0(x0, y0) is a pixel point before image transformation, P(x, y) is a pixel point after image transformation, x0, y0, x and y are spatial coordinates, T is a transformation matrix, wherein a translation transformation matrix is u and v are translation amounts; a rotation transformation matrix is θ is a rotation transformation angle; a scaling transformation matrix is c and d are scaling ratios; the values in the matrix T are adjusted to realize transformation of the pixel point coordinates.

[0014] As a further improvement of the technical scheme of the application, in step S2, when the MSER algorithm extracts feature points of the power pole information board image, the character candidate region is obtained by pixel point sorting, extraction of extreme value regions, determination of maximum stability determination conditions, MSERs region fitting and MSERs region normalization.

[0015] As a further improvement of the technical scheme of the application, when the pixel points are sorted, the image pixel points are subjected to gray scale processing, and the given image is sorted according to the gray scale value.

[0016] As a further improvement of the technical scheme of the application, when the maximum stable determination condition is determined, the extreme value region is taken as a node, a node is constructed into a component tree, and the pixel points R i is an extreme value region, R i ∈D, D is a region, boundary(R i ) is a boundary of the extreme value region, the extreme value region R i is a connected domain of the threshold image I i .

[0017]

[0018] i is a threshold value; I is a gray scale value of the image; the boundary of the component tree defines a containing relationship between the connected domains, when the extreme value region R i is a sub-region of R j , then by moving the component tree upwards, the threshold value i related to the extreme value region is constantly reduced, and the size of the extreme value region R i is constantly increased, when the binary threshold value i changes in [i-Δ, i+Δ], Δ is a small gray scale change, the connected domains are also changed into R i+Δ and R i-Δ , and the region having a minimum change rate q(i) in this change range is considered as an MSERs region.

[0019] As a further improvement of the technical scheme of the application, in step S3, when the non-character region in the character candidate region is removed by using the geometric features, the region width-height ratio, the eccentricity of the ellipse having the same second moment as the region, the proportion of the region pixels in the convex hull, the area of the region divided by the area of the minimum rectangle of the region, the Euler number in the region, and the roundness of the region are processed, a threshold value is set, and the redundant non-character region is deleted.

[0020] As a further improvement of the technical scheme of the application, in step S4, when the single character region is merged into a text line by using the weighted undirected graph, the size of the character region rectangular frame is obtained, the character region rectangular frame is expanded, the overlap rate is calculated, the weighted undirected graph is created by using the overlap rate matrix G, the indexes of the rectangular frames of the connected character regions are connected by the output of the undirected graph, and the adjacent rectangular frames are merged into one rectangular frame.

[0021] Compared with the prior art, the method for detecting and recognizing characters in the electric pole information board image has the following advantages:

[0022] The application takes full advantage of MSER in detecting natural scene text, and gives full play to the advantages of Imbinarize-MSER in multi-scale detection, good candidate region stability and affine invariance. Firstly, the power pole information board is subjected to affine transformation and cutting pretreatment, and combined with the geometric features of characters, the features of the image are accurately described by mathematical morphology, the character detection and recognition of the collected power pole information board are carried out by using the characteristics of fast operation speed, undirected graph merging character region and OCR technology, so as to facilitate the monitoring and management of the operation state of the power transmission line and provide protection for the safety of the power transmission line. BRIEF DESCRIPTION OF DRAWINGS

[0023] The accompanying drawings, which are incorporated herein and constitute part of the specification, illustrate embodiments consistent with the application and, together with the description, serve to explain the principles of the application.

[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced hereinafter. Obviously, for those skilled in the art, other drawings can also be obtained based on these drawings without any creative effort.

[0025] Figure 1 A specific flowchart of the method for character detection and recognition in the power pole information board image. DETAILED DESCRIPTION

[0026] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced hereinafter. Obviously, for those skilled in the art, other drawings can also be obtained based on these drawings without any creative effort.

[0027] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application. However, the present application can be practiced in a variety of ways different from those described herein; obviously, the examples in the specification are only a part of the embodiments of the present application, not all the embodiments.

[0028] The specific embodiments of the present application will be described in detail hereinafter.

[0029] EMBODIMENT

[0030] A method for character detection and recognition in a power pole information board image, comprising the following steps:

[0031] S1: using affine transformation and cutting algorithm to pretreat the collected power pole information board image.

[0032] S1-1: The collected power pole information plate image may be tilted in different directions due to angle or vibration, shooting angle, etc. Therefore, affine transformation is performed first. Affine transformation, also known as affine mapping, specifically refers to a linear transformation of a vector space and a translation to another vector space in geometry. Affine transformation can realize geometric transformations such as translation, rotation, scaling, etc. Affine transformation specifies a 3x3 numerical matrix, each power pole information plate image is regarded as a pixel point arranged in rows and columns, [x y 1] = [x0 y0 1] x T, P0(x0, y0) is the pixel point before image transformation, P(x, y) is the pixel point after image transformation, x0, y0, x, y are spatial coordinates, and T is the transformation matrix, where the translation transformation matrix is u and v are translation amounts; the rotation transformation matrix is θ is the rotation transformation angle; the scaling transformation matrix is c and d are scaling ratios; the values in the matrix T are adjusted to realize the transformation of the pixel point coordinates.

[0033] S1-2: The collected power pole image has much redundant information, large interference and complex background. In order to improve the detection and recognition efficiency of the power pole information plate, a certain size of cutting is performed to reduce the amount of calculation and improve the real-time processing speed.

[0034] S2: The Imbinarize-MSER algorithm is used to extract the feature points of the preprocessed power pole information plate image to obtain the character candidate region.

[0035] S2-1: In order to avoid the poor detection effect of the MSER algorithm when the overall or local contrast is low, the light is sensitive, and the image spot is sensitive, first, the Imbinarize algorithm is used for adaptive threshold processing of the brightness value matrix, and the threshold value is created through the original image. The specific threshold matrix is determined based on the local mean intensity (first-order statistics) of each pixel neighborhood of the image. After adaptive threshold processing, the adverse detection effect caused by low contrast, uneven light, character hole and other factors is reduced, and the accuracy and recall rate of the MSER algorithm are improved.

[0036] S2-2: The MSER algorithm is based on the concept of watershed algorithm, and the image features extracted by the MSER algorithm have good affine invariance. MSER detection uses the concept of watershed in terrain to solve local regions. The watershed algorithm is a set composed of a series of regions, and the key depends on the threshold value when the region is fused, but these threshold values are unstable and do not exist after fusion. The MSER algorithm is to find the basin or plateau in the region, that is, when the water level changes within a certain range, the area of the region changes very slowly.

[0037] Image I is a mapping of a region D to gray scale S: D→S. S is orderable, and in gray scale image S = {0, 1, … 255}. Neighborhood relation If A adopts four-neighborhood, for two pixel points p, q∈D, they are neighborhood relation if and only if The algorithm mainly consists of four steps, as follows:

[0038] (1) Pixel point ordering

[0039] The pixel points of the image are processed by gray scale, and the given image is ordered according to the gray scale value, also known as box ordering. It is assumed that the input elements to be sorted are equally likely to fall into equal interval value intervals. The equal interval interval is called a box, and each box saves the elements falling in the interval. Each box is automatically sorted, and finally these boxes are concatenated.

[0040] (2) Extraction of extreme value region

[0041] The extraction of the extreme value region is to calculate the extreme value region on each threshold image in the order from large to small according to the threshold gray scale value. This process finally generates a tree-like data structure. The data structure becomes a component tree, and each layer of the component tree corresponds to a threshold image. Each node on the layer represents an extreme value region responding to the threshold image. In order to improve the calculation efficiency and quickly extract the MSER, the optimal binary tree separation set forest algorithm is used to establish the component tree, and the optimal binary tree and the separation set forest data structure are combined. Each pixel point in the image is allocated a node, and according to the size ordering result of the pixel points, they are placed into the optimal binary tree one by one. The specific order is that each pixel point corresponds to a node index number.

[0042] (3) Maximum stable determination condition

[0043] In determining the maximum stable determination condition, the extreme value region is taken as the node, and the node is constructed into a component tree. The pixel point R i is the extreme value region, R i ∈D, D is the region, boundary(R i ) is the extreme value region boundary, and the extreme value region R i is the connected domain of the threshold image I i :

[0044]

[0045] i is the threshold; I is the gray scale value of the image, and the extreme value region R i is a pixel set, and the pixels in the pixel set are in four-neighborhood relation; the boundary of the component tree defines the inclusion relation between the connected domains, and when the extreme value region R i is R jIf the sub-region of R By moving up the component tree, the threshold i related to the extreme region R i is constantly decreasing, while the size of the extreme region R i+Δ is constantly increasing, when the binarization threshold i varies in [i-Δ, i+Δ], Δ is a small gray level change, the connected domain also changes to R i-Δ correspondingly. The region with a small change rate q(i) in this change range is considered as an MSERs region.

[0046] In order to ensure the accuracy of the stable region, the threshold takes a dynamic threshold algorithm for iteration, and after several iterations, the region with little area change is found and determined as a stable region.

[0047] (4) MSERs region fitting

[0048] The maximum stable extreme region is a region with any irregular shape, and it is difficult to describe the characteristics of the irregular shape. In order to better describe the region, the maximum stable extreme region is represented as an ellipse by using the method of constructing affine invariants. From a mathematical point of view, covariance is an extension of variance, which measures the change trend of two feature vectors in the same direction. From a geometric point of view, covariance corresponds to the elliptical distribution of features, and the larger the covariance, the more the corresponding ellipse is flat, that is, the feature vectors are clustered around the principal axis. The second order central distance is a generalized covariance. Therefore, the second order central distance based on vector representation is constructed by using the pixel information of the irregular maximum stable extreme region, and the second order central distance is used for adjusting the irregular maximum stable extreme region.

[0049] (5) MSERs region normalization

[0050] After the MSERs region is fitted as an ellipse, in order to construct a feature descriptor with affine invariance, the ellipse fitting region needs to be expanded to a measurement area, and then the measurement area is normalized according to the specified radius. The feature descriptor is extracted on the normalized image.

[0051] Normalization mainly needs three steps: first, affine normalization is performed on the measurement area corresponding to the MSERs fitting area.

[0052] Second, gradient histogram statistics are performed in the affine normalized area, the maximum value of the histogram is found, and the main direction of the normalized gradient corresponding to the maximum value is found.

[0053] Third, the normalized area image is rotated according to the main direction, and the specific rotation normalization formula is: l is the circumference of the ellipse, l max is the maximum value of the circumference of the ellipse, l minis the minimum value of the ellipse perimeter. In this way, the detected MSERs ellipse feature measure region normalization has the local image region of affine invariance.

[0054] A large number of character candidate regions are obtained by MSER detection algorithm.

[0055] S3: Geometric features and mathematical morphology are used to remove non-character regions in the character candidate region, and single character regions are obtained.

[0056] Most of the characters in the power pole information board can be extracted by the MSER algorithm, but it may also detect many stable regions of non-character regions, i.e. pseudo-character regions. This is because the background region and the character region have similar features, and geometric features and morphological algorithms are needed to filter non-character regions.

[0057] 1) On the basis of the features extracted by the MSER algorithm, the non-character regions are reduced by extracting the geometric features of the binary image. Mainly use the region width-height ratio, the eccentricity of the ellipse with the same second moment of the region, the convex hull region pixel ratio, the area of the region divided by the area of the smallest rectangle containing the region, the Euler number in the region, and the region roundness. According to the actual situation, set the threshold value, and delete the redundant non-character regions.

[0058] a. Width-height ratio calculation method: W / H, W is the width of the region rectangular frame, and H is the height of the region rectangular frame. The width-height ratio generally takes the value of 3-5.

[0059] b. Calculation method of eccentricity of ellipse: jj / cz, jj is the focal length of ellipse, and cz is the value of the major axis of ellipse. The value is between 0 and 1. (0 and 1 are special cases, the eccentricity of ellipse is 0, and the eccentricity of ellipse is 1. ) The eccentricity of ellipse generally takes the value of 0.95-0.98.

[0060] d. Calculation method of convex hull pixel ratio: Area / ConArea, Area is the total number of pixel points in the region, and ConArea is the number of pixel points in the convex hull image. The convex hull pixel ratio generally takes the value of 0.3-0.4.

[0061] e. Calculation method of area of region divided by area of smallest rectangle containing region: Area / bm, Area is the total number of pixel points in the region, and bm is the area of the region bounding box. Generally, 0.3

[0062] f. Calculation method of Euler number: ds-ks, ds is the number of region objects, and ks is the number of holes in the object. Generally, (ds-ks) <-4.

[0063] g. Calculation method of region roundness: (4 x Area x pi) / Per2 Area is the total number of pixels in the region, Per is the perimeter of the region, and pi is π. For a circle, the roundness value is 1. Generally, the value is (4 x Area x pi) / Per 2 >4.

[0064] For pixels that do not meet the geometric characteristics, they will be removed.

[0065] 2) In order to further remove the redundant non-character area, it is necessary to use mathematical morphology skeleton operation for correction. The skeleton concept was first proposed by Blum and is called the medial axis transform. The skeleton is a thin curve that expresses the shape of an object in accordance with the connectivity and topological structure of the original shape. Skeletonization is a very important transform in image analysis and shape description, and the skeleton is a topological description of the geometric morphology of the image. If an image is used to illustrate the meaning of the skeleton, it is to imagine that at the beginning, the boundaries of the target are ignited at the same time, the flames spread to the interior of the target at a uniform speed, and the flames are extinguished when the leading edges of the flames intersect. The set of flame extinguishing points constitutes the medial axis, which is the skeleton of the image. Through skeleton operation, the internal pixels will be removed. As the medial axis of the image, the skeleton is specifically implemented by using an index table method, which needs to calculate the value of the connected domain around the pixel point, check the corresponding item in the index table to determine whether to keep the point, and generally the internal points, boundary points and breakpoints are not deleted, and other useless points will be removed.

[0066] 3) In order to avoid the influence of removing pixels on the original target characteristics, after the skeleton operation, the filling area processing is performed to correct the image. The closing operation is a commonly used operation in mathematical morphology, which can fill relatively narrow discontinuities, long and narrow gullies and small holes, and can also fill small cracks and contour line breaks in the region. The closing operation specifically uses a diamond structure element, which is close to the shape of the character. After the closing operation, the opening operation is performed. The opening operation and the closing operation are dual, which effectively removes small noise, eliminates character burrs, removes small bright details, and makes the character region smoother.

[0067] 4) The distance transform needs to be performed on the binary image, and the Euclidean distance is specifically used. In two-dimensional space, the Euclidean distance between two points (x1, y1) and (x2, y2) in the image is: The distance value of this point is marked by calculating the distance between the foreground pixel point of the skeleton image and the region boundary corresponding to this pixel point. The skeleton distance image is obtained by traversing the entire image. Then the skeleton feature value is calculated. The calculation method is: jlfc / jljz, jlfc is the distance variance, and jljz is the distance average value. At the same time, a threshold value is set, and the values greater than the threshold value are filtered out to further remove the non-character area. The threshold value is generally 0.4-0.7.

[0068] S4: Merge single character regions into text lines using undirected graph.

[0069] After non-character region filtering, the remaining is basically single character region, in order to make the Chinese characters, letters and numbers in the power pole information board express complete semantic information, it is necessary to merge single character region into text line.

[0070] 1) Get the size of the character region rectangle box, the specific format is [x y w h], (x, y) is the coordinate of the upper left corner of the rectangle box, w and h are the width and height of the rectangle box respectively.

[0071] 2) Expand the size of the character region rectangle box by a certain amount, the expansion method: x×(1-n), y×(1-n), w×(1+n), h×(1+n), n is generally taken as 0.02-0.07.

[0072] 3) Calculate the overlap rate, the calculation method: Area is the total number of pixels in the region, region1 and region2 are two regions. Set the overlap rate of the region boundary rectangle box to zero, which can simplify the graph representation.

[0073] 4) Create a weighted undirected graph using the overlap rate matrix G, each non-zero item in the matrix G specifies an edge of the graph, and the weight of the edge is equal to the value of the item. Find the image connected domain through the undirected graph, each node in the graph corresponds to each pixel in the image, each edge connects a pair of adjacent pixels, and the weight of the edge represents the non-negative similarity between adjacent pixels in terms of gray scale, color or texture. A segmentation of the image is a cut of the graph, and each region divided is corresponding to a subgraph in the graph. The algorithm of segmentation uses Grarbcut algorithm based on Gaussian mixture model, which has less iteration. The optimal segmentation is to make the similarity in the divided subgraph maximum and the similarity between the subgraphs minimum.

[0074] 5) The output of the undirected graph connects the indexes of the rectangle boxes of the character regions, the minimum and maximum values of the single rectangle box are calculated to eliminate isolated character regions, and then the adjacent rectangle boxes are merged into a rectangle box using the indexes to suppress false character detection and recognition.

[0075] S5: Use OCR technology to recognize the text line.

[0076] The OCR (Optical Character Recognition) technology is a practical technology that the computer automatically and rapidly identifies the text on paper and converts it into editable text. It is an important part of a new generation of computer intelligent interface and an important branch of pattern recognition field. First, the optical character recognition language data file package is installed, and then the character recognition is performed on the merged character region, and the recognition rate is above 97%.

[0077] The above merely describes the specific embodiments of the present application, enabling those skilled in the art to understand or implement the present application. Although the foregoing embodiments are described in detail, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can be modified, or some or all of the technical features can be replaced equivalently, and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments, and they should be covered in the protection scope of the claims.

Claims

1. A method for character detection and recognition in power pole sign image, characterized in that, It comprises the following steps: S1: using affine transformation and clipping algorithm to preprocess the collected power pole information board image; S2: using Imbinarize-MSER algorithm to extract the feature points of the preprocessed power pole information board image, and obtaining the character candidate region; using Imbinarize algorithm to perform adaptive threshold processing on the brightness value matrix, and the threshold value is created by the original image, and the specific threshold matrix is determined based on the local mean intensity of each pixel neighborhood of the image; S3: using geometric features and mathematical morphology to remove non-character regions in the character candidate region, and obtaining single character regions; in step S3, when using geometric features to remove non-character regions in the character candidate region, the region width-height ratio, the eccentricity of the ellipse with the same second moment as the region, the proportion of the region pixels in the convex hull, the area of the region divided by the area of the smallest rectangle of the region, the Euler number in the region, and the roundness of the region are used for processing, the threshold value is set, and the redundant non-character regions are deleted; the mathematical morphology skeleton operation is used for correction; after the skeleton operation, the region filling processing is performed to correct the image; the binary image is subjected to distance transformation, and the Euclidean distance is used specifically; S4: using an undirected graph to merge single character regions into text lines; in step S4, when using the weighted undirected graph to merge single character regions into text lines, the size of the character region rectangular frame is obtained, the character region rectangular frame is expanded, the overlap rate is calculated, the weighted undirected graph is created using the overlap rate matrix G, the indexes of the rectangular frames of the character regions are connected through the output of the undirected graph, and adjacent rectangular frames are merged into one rectangular frame; S5: using OCR technology to recognize the text line.

2. The method for character detection and recognition in images of power pole signs according to claim 1, characterized in that, In step S1, the affine transformation specifies a 3x3 numerical matrix, each power pole information plate image is regarded as a pixel point arranged in rows and columns, , is a pixel point before image transformation, is a pixel point after image transformation, , , , is a spatial coordinate, is a transformation matrix, wherein the translation transformation matrix , , is a translation amount; the rotation transformation matrix , is a rotation transformation angle; the scaling transformation matrix , , is a scaling ratio; the adjustment matrix The value in the adjustment matrix is used to realize the transformation of the pixel point coordinate.

3. The method of claim 1, wherein, In step S2, when the MSER algorithm extracts the feature points of the power pole information board image, the pixel point sorting, the extreme value region extraction, the maximum stability determination condition determination, the MSERs region fitting, and the MSERs region normalization are used to obtain the character candidate region; The MSERs region normalization comprises the following steps: Firstly, the measurement area corresponding to the MSERs fitting area is subjected to affine normalization; Secondly, the gradient histogram statistics are performed in the affine normalized area, the maximum value of the histogram is found out, and the main direction of the normalized gradient corresponding to the maximum value is normalized; Third, according to the main direction of the normalized region image is rotated again normalized, specific rotation normalization formula: , is the ellipse perimeter, is the ellipse perimeter maximum, is the ellipse perimeter minimum; detected MSERs ellipse shape feature measurement region normalized local image region has affine invariance.

4. The method of claim 3, wherein, When the pixel point sorting is performed, the image pixel points are subjected to gray processing, and the given image is sorted according to the gray value size.

5. The method of claim 3, wherein the method further comprises: In determining the maximum stable decision condition, the extreme value region is taken as a node, a node is constructed into a tree, and the pixel points , , are extreme value regions, , are regions, are extreme value region boundaries, and the extreme value region is a connected domain of the threshold image . ； is a threshold value; is a gray value of the image; the boundary of the component tree defines the inclusion relation between connected components, when the extreme value region is a sub-region of , then ; by moving the component tree upward, the threshold value associated with the extreme value region is constantly reduced, while the size of the extreme value region is constantly increased, when the threshold value of the binaryzation changes , is a small gray value change, the connected components also change and correspondingly, the region with a small change rate in this change range is considered as an MSER region.

Images