CN110569848A

CN110569848A - feature extraction method and system for power equipment nameplate

Info

Publication number: CN110569848A
Application number: CN201910736623.8A
Authority: CN
Inventors: 吴彦直; 李林; 郑志曜; 高一波
Original assignee: Individual
Current assignee: Individual
Priority date: 2019-08-09
Filing date: 2019-08-09
Publication date: 2019-12-13

Abstract

the invention provides a method and a system for extracting characteristics of a nameplate of power equipment, wherein the method and the system are suitable for extracting characteristics of a nameplate image of the power equipment; traversing each character area, and performing character recognition on each character area in the traversing process; and finally, based on the original image and the character recognition result corresponding to each character area, using a sift algorithm to realize the feature extraction of the nameplate image after the position of the nameplate feature point is determined. Compared with the traditional content-based feature extraction method, the method has the advantages that stable feature points can be extracted from images with different illumination, rotation and scale change by the features, and the storage structure combining the feature index and the feature point queue storage is provided, so that the method has better matching capability on feature similar points in local areas of the images.

Description

feature extraction method and system for power equipment nameplate

Technical Field

The invention relates to the field of computer vision, in particular to a method and a system for extracting the characteristics of a nameplate of power equipment based on a character region.

background

As a special image, the power equipment nameplate image is relatively single in texture and color information and is difficult to control under the conditions of illumination and the like for general images, so that the requirement for feature extraction of the power equipment nameplate image is more severe, and the proper power equipment nameplate feature extraction method plays an important role in the whole nameplate identification technology.

the traditional nameplate feature extraction method based on content mainly aims at the relevant research of nameplate shape features, and is mainly divided into a feature extraction method based on a region and a feature extraction method based on an angular point according to shape description.

most of the conventional feature extraction methods are simple in implementation principle, are very sensitive to specific attributes in an image, such as scale and illumination change, and cannot extract stable feature information, so that the problems of poor robustness, low reliability and the like exist.

disclosure of Invention

the invention aims to solve the technical problems of poor robustness and low reliability in the prior art, and provides a method and a system for extracting the nameplate characteristics of electric power equipment based on a character region.

the technical scheme adopted by the invention for solving the technical problems is as follows: a feature extraction method for constructing a nameplate of power equipment comprises the following steps:

s1: acquiring a nameplate image of the power equipment;

s2: after the nameplate image collected in the step S1 is subjected to noise reduction processing, region cutting is performed according to the position of the nameplate, and then the position region of the nameplate is obtained;

S3: aiming at the position area obtained by the processing in the step S2 and the position of the recorded characters on the nameplate, carrying out area segmentation on the position area by using an MSER algorithm to obtain a plurality of character areas;

S4: traversing each character area obtained in the step S3, and performing character recognition on each character area in the traversing process; the recognition result of each character area is used as an index label to further calibrate each character area;

S5: and (4) identifying the original image corresponding to each divided character area obtained in the step (S3) and the identification result obtained in the step (S4) by using a sift algorithm, and then realizing the feature extraction of the nameplate image after determining the position of the nameplate feature point.

further, the image preprocessing in step S2 specifically includes the following steps:

S21: graying the nameplate image collected in the step S1 to obtain a grayed image;

s22: performing noise reduction processing on the gray image obtained in the step S21 by using a gaussian filter to obtain a noise-reduced image;

s23: and (5) performing edge detection on the noise-reduced image obtained in the step (S22) by using a canny operator, and performing region cutting on the position of the nameplate to obtain the position region of the nameplate, specifically, cutting and filtering the background region in the image.

Further, the step of performing region segmentation on the nameplate image in step S3 includes:

S31, decomposing the nameplate image into a plurality of different connected domains by using a MSER algorithm;

and S32, traversing each communication domain obtained by decomposition in the step S31 by using an NMS algorithm, filtering and merging related communication domains, and taking an area obtained by merging related communication domains as a character area.

further, in step S4, the character region obtained by dividing in step S3 is subjected to character recognition using an OCR character recognition algorithm, wherein the recognition result is further used as an index label of the character region.

Further, the step of extracting the feature of the nameplate image by using the sift algorithm in the step S5 includes:

s51: establishing a DoG multi-scale pyramid model based on the acquired nameplate image, and further obtaining the positions of all feature points in the image by removing unstable and marginalized feature points in the image after calculating the local extreme value of the image by using the DoG multi-scale pyramid model;

s52: calculating a feature point descriptor according to the positions of all the feature points acquired in the step S51;

s53: and generating image features of the nameplate based on the feature descriptors.

Further, the step of calculating the feature point descriptor in step S52 includes:

S521, acquiring all the feature points in the step S51, taking a pixel point set with the neighborhood scale of the feature point being m as a processing object, and calculating the pixel gradient, the gradient amplitude and the direction of the feature points;

s522, performing feature point adjacent region, and taking n multiplied by n sub-regions obtained by dividing as statistical units;

S523: counting gradient information in r directions in each region, and taking n multiplied by n sub-regions divided in the step S522 as a counting unit according to a calculation method of the pixel gradient, the gradient amplitude and the direction of S521, and counting the pixels, the gradient, the amplitude and the direction of k multiplied by k neighborhood to obtain k histogram statistical graphs; wherein, k is n × n, and the k histogram statistics are uniformly mapped into n × n × r dimensional gradient information;

S524: combining the statistical information of each region to obtain n multiplied by r dimensional characteristic vectors, and calculating to obtain a descriptor l by utilizing a normalization algorithm based on the obtained characteristic vectors, wherein the calculation formula of the normalization algorithm is as follows:

wherein l ═ l₁,l₂,...,l_t]。

further, after the feature extraction of the nameplate is completed, the recognition result of the step S4 is stored as an index tag for each character region obtained in the step S3 for the first time, and the feature points extracted in the step S5 are stored for the second time.

the invention discloses a characteristic extraction system of a power equipment nameplate, which comprises the following modules:

the data acquisition module is used for acquiring a nameplate image of the power equipment;

the position area acquisition module is used for carrying out noise reduction treatment on the acquired nameplate image and carrying out area cutting on the position of the nameplate so as to obtain the position area of the nameplate;

The character region acquisition module is used for performing region segmentation on the position region by using a MSER algorithm according to the position region obtained by processing and the position of the recorded characters on the nameplate to obtain a plurality of character regions;

The character recognition module is used for traversing each acquired character area and recognizing characters of each character area in the traversing process; the recognition result of each character area is used as an index label to further calibrate each character area;

and the feature extraction module is used for extracting the features of the nameplate image after determining the positions of the nameplate feature points by using a sift algorithm for the original image and the character recognition result corresponding to each segmented character area.

further, the feature extraction module further comprises the following sub-modules:

the feature point position determining module is used for establishing a DoG multi-scale pyramid model based on the acquired nameplate image, and further obtaining the positions of all feature points in the image by removing unstable and marginal feature points in the image after calculating the local extreme value of the image by using the DoG multi-scale pyramid model;

The characteristic point descriptor calculation module is used for calculating the characteristic point descriptors according to the positions of all the characteristic points acquired by the characteristic point position determination module;

And the image feature generation module is used for generating the image feature of the nameplate based on the feature descriptor calculated by the feature point descriptor calculation module.

Further, the feature point descriptor computation module comprises the following sub-modules:

The data preprocessing module is used for acquiring all the characteristic points determined in the characteristic point position determining module, taking a pixel point set with the neighborhood scale size of the characteristic point being m as a processing object, and calculating the pixel gradient, the gradient amplitude and the direction of the characteristic point;

The statistical unit dividing module is used for dividing the adjacent region of the feature points, and taking the n multiplied by n sub-regions obtained by dividing as statistical units;

the data statistics module is used for counting gradient information of r directions in each region, taking n multiplied by n sub-regions divided by the statistics unit division module as a statistics unit according to a calculation method of the data preprocessing module on pixel gradient, gradient amplitude and direction, and counting pixels, gradient, amplitude and direction of k multiplied by k neighborhood to obtain k histogram statistics graphs; wherein, k is n × n, and the k histogram statistics are uniformly mapped into n × n × r dimensional gradient information;

and the descriptor calculation module is used for combining the statistical information of each region to obtain n multiplied by r dimensional feature vectors, and calculating to obtain the descriptor l by utilizing a normalization algorithm based on the obtained feature vectors.

in the method and the system for extracting the characteristic of the nameplate of the power equipment based on the character region, stable characteristic points are extracted from images with different illumination, rotation and scale change, and a storage structure combining characteristic indexes and characteristic point queue storage is provided, so that the method and the system have better matching capability on characteristic similar points in local regions of the images.

The invention provides a character region-based electric power equipment nameplate feature extraction method and system, provides a storage structure combining feature indexes and feature point queue storage, and provides stable data support for subsequent feature matching.

Drawings

the invention will be further described with reference to the accompanying drawings and examples, in which:

FIG. 1 is a flow chart of a method for extracting characteristics of a nameplate of an electrical device according to the present disclosure;

FIG. 2 is a view of the location area of the tag;

FIG. 3 is a pictorial illustration of a text region of the tag;

FIG. 4 is a flow chart of regional feature point acquisition;

fig. 5 is a structural diagram of a feature extraction system of a nameplate of an electrical device disclosed by the invention.

Detailed Description

for a more clear understanding of the technical features, objects and effects of the present invention, embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

The invention provides a character region-based electric power equipment nameplate feature extraction method and system, which are most suitable for feature extraction of an electric power equipment nameplate image. Compared with the traditional content-based feature extraction method, the method has the advantages that stable feature points can be extracted from images with different illumination, rotation and scale changes by means of the features, a storage structure combining feature indexes and feature point queue storage is provided, and the matching capability of feature similar points in local regions of the images is good.

referring to fig. 1, which is a flowchart of a method for extracting features of a nameplate of an electrical device disclosed in the present invention, the method for extracting features of a nameplate of an electrical device disclosed in the present invention specifically includes the following steps:

S1, collecting images, wherein in the embodiment, an RGB camera is used for collecting nameplate images of the transformer aiming at the transformer;

s2, preprocessing the image, and determining and cutting a nameplate position area; the method specifically comprises the following steps:

Firstly, carrying out graying processing on the nameplate image collected in the step S1 to obtain a grayed image;

secondly, performing noise reduction processing on the processed gray images by using a Gaussian filter to obtain noise-reduced images;

And finally: the method includes the steps of performing edge detection on an image subjected to graying and noise reduction by using a canny operator, performing region cutting on the image according to the position of a nameplate, and further obtaining a position region of the nameplate, specifically, performing cutting and filtering on a background region in the image, and referring to fig. 2 for specific implementation effects, wherein the specific implementation effects are obtained by filtering the background region of the nameplate.

s3, carrying out region segmentation on the nameplate image obtained in the step S2 to obtain a character region; the method specifically comprises the following steps:

aiming at the position area obtained by the processing of the step S2 and the position of the recorded characters on the nameplate, carrying out area segmentation on the position area by using an MSER algorithm to obtain a plurality of different connected domains; and finally, traversing each connected domain obtained by current decomposition by using an NMS algorithm, filtering and merging related connected domains, and taking the region obtained by merging related connected domains as a character region. Referring to fig. 3, the area defined by the rectangular or square frame is the connected domain obtained by dividing, and the connected domain adjacent to or having the text link pipe is merged to obtain the complete area which is the text area.

S4, acquiring an area index label; in this embodiment, the character areas obtained by dividing in step S3 are subjected to character recognition by using an OCR character recognition algorithm, wherein the obtained character recognition result is further used as an index label of the character area.

S5, obtaining region feature points, and extracting the feature points of the original image corresponding to the segmented character region obtained in the S3 by using a sift algorithm; the step of extracting the feature of the nameplate image by using the sift algorithm includes (for a specific flow implementation diagram, please refer to fig. 4):

s52: calculating a feature point descriptor according to the positions of all the feature points acquired in the step S51; wherein the step of calculating the descriptor comprises:

s521: acquiring all feature points in S51, taking a pixel point set with a feature point neighborhood scale of 3 sigma as a processing object, and calculating to obtain the pixel gradient, the gradient amplitude and the direction of the feature points aiming at the processing object;

S522: dividing the feature point adjacent region into 16 sub-regions of 4 x 4, and taking the sub-regions as statistical units;

s523: creating a gradient direction histogram, wherein for the 360-degree histogram, 10 degrees are taken as a statistical unit, the histogram is divided into 36 columns, and for each column, the gradient amplitude and the gradient direction of each regional pixel point set are calculated;

s524: and taking the region direction with the maximum gradient amplitude as a main direction, and taking the region direction larger than 80% of the main direction as an auxiliary direction to determine the gradient direction of the region.

s525: and (3) counting gradient information in 8 directions in each region, counting the gradient, the amplitude and the direction of the 16 × 16 neighborhood pixels by taking the sub-regions divided in the step S522 as a counting unit according to the calculation method of the pixel gradient, the gradient amplitude and the direction of S521, and counting the result by taking a certain sub-region as a unit to obtain 16 histogram statistical graphs. And finally, uniformly mapping the 16 histogram statistical graphs into 4 multiplied by 8-128-dimensional gradient information, and recording the gradient information as H-H₁,h₂,……,h₁₂₈。

s526: and combining the statistical information of each region to obtain a 128-dimensional feature vector, and calculating by adopting a normalization calculation method aiming at the obtained 128-dimensional feature vector to obtain a descriptor l:

wherein l ═ l₁,l₂,...,l_t]。

s6, storing character region characteristics based on the data obtained by processing in the steps S3-S4, wherein the storage mode is shown in Table 1, and the storage is convenient for later data calling and obtaining; for each character region (number 1, 2, 3, …, n) acquired in step S3, the recognition result in step S4 is stored once as an index tag (number 1, 2, … …, n), and the recognition result in step S5 is stored twice as a feature point (number 1, 2, … …, n).

TABLE 1

please refer to fig. 5, which is a feature extraction system for a nameplate of an electrical device according to the present disclosure, as shown in the figure, the system includes a data obtaining module L1, a location area obtaining module L2, a text area obtaining module L3, a text recognition module L4, and a feature extraction module L5, wherein:

the data acquisition module L1 is used for acquiring a nameplate image of the power equipment;

the position area obtaining module L2 is configured to perform noise reduction on the collected nameplate image, and perform area cutting on the location of the nameplate, so as to obtain a position area of the nameplate;

the character region acquisition module L3 is used for performing region segmentation on the position region by using MSER algorithm according to the position region obtained by processing and the position of the recorded characters on the nameplate to obtain a plurality of character regions;

the character recognition module L4 is configured to traverse each acquired character region, and perform character recognition on each character region in the traversing process; the recognition result of each character area is used as an index label to further calibrate each character area;

the feature extraction module L5 is used for extracting features of the nameplate image after determining the position of the nameplate feature point by using a sift algorithm for the original image and the character recognition result corresponding to each segmented character region; the feature extraction module further comprises a feature point position determination module L51, a feature point descriptor computation module L52 and an image feature generation module L53, wherein:

the feature point position determining module L51 is used for establishing a DoG multi-scale pyramid model based on the acquired nameplate image, and further obtaining the positions of all feature points in the image by removing unstable and marginal feature points in the image after calculating the local extreme value of the image by using the DoG multi-scale pyramid model;

The characteristic point descriptor calculation module L52 is used for calculating a characteristic point descriptor according to the positions of all the characteristic points acquired by the characteristic point position determination module; the feature point descriptor computation module further includes a data preprocessing module L521, a statistical unit dividing module L52, a data statistics module L523, a descriptor computation module L524, and an image feature generation module L525, where:

the data preprocessing module L521 is configured to acquire all feature points determined in the feature point position determining module, use a set of pixel points with a neighborhood scale size of m of the feature point as a processing object, and calculate a pixel gradient, a gradient amplitude and a direction of the feature point;

the statistical unit dividing module L522 is configured to perform processing on the feature point neighboring region, and use the n × n sub-regions obtained by dividing as statistical units;

the data statistics module L523 is configured to count gradient information in r directions in each region, and according to a calculation method of the data preprocessing module on pixel gradients, gradient amplitudes and directions, take n × n sub-regions divided by the statistics unit division module as a statistics unit, count pixels, gradients, amplitudes and directions of k × k neighborhoods, and obtain k histogram statistics diagrams; wherein, k is n × n, and the k histogram statistics are uniformly mapped into n × n × r dimensional gradient information;

The descriptor computation module L524 is configured to merge statistical information of each region to obtain an n × n × r-dimensional feature vector, and compute a descriptor L by using a normalization algorithm based on the obtained feature vector.

the image feature generation module L525 is configured to generate an image feature of the nameplate based on the feature descriptor calculated by the feature descriptor calculation module.

While the present invention has been described with reference to the embodiments shown in the drawings, the present invention is not limited to the embodiments, which are illustrative and not restrictive, and it will be apparent to those skilled in the art that various changes and modifications can be made therein without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. A feature extraction method for a power equipment nameplate is characterized by comprising the following steps:

S1: acquiring a nameplate image of the power equipment;

s3: aiming at the position area obtained by the processing of the step S2, based on the position of the recorded characters on the nameplate, the MSER algorithm is used for carrying out area segmentation on the position area to obtain a plurality of character areas;

S4: traversing each character area obtained in the step S3, and performing character recognition on characters recorded in each character area in the traversing process; the recognition result of each character area is used as an index label to further calibrate each character area;

s5: and combining the original image corresponding to the character area, namely the nameplate image of the power equipment obtained in the step S1, and after the position of the nameplate feature point is determined by using a sift algorithm, realizing feature extraction of the nameplate image, wherein the extracted data is used for the later research on nameplate identification.

2. The feature extraction method according to claim 1, wherein the image preprocessing in step S2 specifically includes the following steps:

S23: and (3) aiming at the noise-reduced image obtained in the step (S22), carrying out edge detection on the noise-reduced image by using a canny operator, and carrying out region cutting aiming at the position of the nameplate so as to obtain the position region of the nameplate, wherein the region collection refers to cutting and filtering aiming at the background region in the image.

3. The feature extraction method according to claim 1, wherein the step of performing region segmentation on the nameplate image in step S3 includes:

4. the feature extraction method of claim 1, wherein step S4 performs character recognition on the character region obtained by segmentation in step S3 by using an OCR character recognition algorithm, wherein the recognition result is further used as an index tag of the character region.

5. the feature extraction method according to claim 1, wherein the feature extraction of the nameplate image using a sift algorithm in step S5 comprises:

6. the feature extraction method according to claim 5, wherein the step of calculating the feature point descriptor in step S52 includes:

s522, performing region division on the adjacent region of each feature point, and taking n multiplied by n sub-regions obtained by division as statistical units;

S523: counting gradient information of r directions in each region, taking n multiplied by n sub-regions divided in step S522 as a counting unit according to a calculation method of pixel gradient, gradient amplitude and direction of step S521, and counting pixels, gradient, amplitude and direction of k multiplied by k neighborhood to obtain k histogram statistical graphs; wherein, k is n × n, and the k histogram statistics are uniformly mapped into n × n × r dimensional gradient information;

wherein l ═ l₁,l₂,...,l_t]。

7. the feature extraction method according to claim 1, wherein after the feature extraction of the name plate is completed, the recognition result of the step S4 is stored as an index tag for one time and the feature point extracted in the step S5 is stored for the second time for each character region obtained in the step S3.

8. the feature extraction system for the nameplate of the power equipment is characterized by comprising the following modules:

9. the feature extraction system of claim 8, wherein the feature extraction module further comprises the following sub-modules:

10. the feature extraction system of claim 9, wherein the feature point descriptor computation module comprises the following sub-modules: