RU190639U1 - SYSTEM OF GENERATION OF IMAGES CONTAINING TEXT - Google Patents

Abstract

The utility model relates to the field of data processing. The technical result of the utility model is the creation of a system for generating images containing text with less complexity and time of operation, due to the rejection of the use of images obtained during the filming process, in favor of images generated using electronic computing means of image generation. 3 hp f-ly, 4 ill.

Description

The utility model relates to the field of data processing, in particular, to systems for generating images containing text, and can be used in optical character recognition systems for symbolic information, including texts located on labels and plates of a predetermined format.

Optical text recognition is used in many areas: to translate scanned copies of documents into electronic form, to identify vehicles according to the state license plate image, to fix and translate any text found in the image stream of optical devices into a text view.

Optical text recognition methods can be divided into two categories: methods based on heuristic algorithms for analyzing the structure of the text, followed by highlighting text information, and methods based on machine learning algorithms, whose parameters are selected at a preliminary stage in the process of analyzing a large number of images marked on them manual text operators.

Heuristic algorithms, as a rule, have high speed, because they are based on a rapid analysis of the characteristic frequencies of background images and text fonts. However, the disadvantages of such algorithms include the high degree of attachment of algorithmic blocks to a specific format of recognized texts, and, accordingly, the greater complexity of adapting these algorithms to new formats of backgrounds and text fonts.

Algorithms based on machine learning do not have the above disadvantage: to introduce a new text format that needs to be recognized, as a rule, it is enough to form a training sample of the required volume with the manually specified position of the text symbols. As a rule, neural networks have recently been used as a recognition algorithm; therefore, it is enough to indicate new neural network architectures and other neuron weights to go to new text formats, labels, plates, and the recognition algorithm itself does not require significant modifications. A significant lack of algorithms based on machine learning should be attributed to the large amount of input data required for both learning the algorithm and testing it. The preparation of this data is the most time-consuming part of the learning process of optical recognition algorithms for text images and characters.

Given that OCR systems are aimed at a specific format, it is possible to create systems and methods for generating images of texts in exactly the format for which the original OCR system is intended. Moreover, when generating images of texts, it is definitely possible to restore the position of each letter, so there is no need to manually perform a time-consuming operation of marking the position of characters in the image, therefore, the process of forming the training sample in this case occurs several orders of magnitude faster.

The closest to the claimed utility model is an image generation system (patent CN106874937), based on the use of a vector font and a binarized image of the target word, in which images are generated by changing the chromaticity, background and other parameters of the original images obtained as a result of shooting. This system is selected as a prototype of the claimed utility model.

The disadvantage of the prototype system is that it generates images by modifying the original images obtained as a result of shooting, and not generated artificially using electronic computing means, so it requires a large number of source images taken, which are time-consuming to find. . This increases the complexity and time of operation of this system, spent on searching or creating the original captured images.

The technical result of the utility model is the creation of a system for generating images containing text with less complexity and time of operation, due to the rejection of the use of images obtained during the filming process, in favor of images generated using electronic computing means of image generation.

The technical result is achieved by creating a system for generating images containing text containing a text imaging unit 1 and a background imaging unit 2, the outputs of which are connected to the input of the seamless insert unit 3, the input of which is connected to the input of the final image variation block 4,

- block 2 forming the background image made with the possibility of forming the image of the background and transfer it to the block seamless insert,

- block 1 of the imaging text is made with the possibility of forming an image of the text and transfer it to

- block 3 seamless insert, which is made with the possibility of forming the final image by overlaying the image of text on the background image, while reducing image defects on the border of the text image and the background image, as well as the ability to transfer the final image to

- block 4 variations of the final image, which is made with the possibility of modifying the final image using image variation methods used to train computer vision algorithms, as well as the ability to transfer the modified final image to the external block 5 learning recognition algorithms, which is part of the external image recognition system .

In a preferred embodiment of the system, the background image forming unit 2 comprises a background image selection unit 7, the input of which is connected to a pre-formed background image base 6, and the output of which is connected to the input of the scale selection section 8 and the background image area,

- the block of the choice of 7 background images made with the possibility of a random selection of the background image of the pre-formed base 6 background images and transfer it

- in block 8 of the choice of scale and the background image area, which is made with the possibility of a probable change of the background image scale and a random selection of the background image area of a given size as the final background image for the final image, as well as the ability to transfer the background image into the seamless insert unit 3.

In the preferred embodiment of the system, the text imaging unit 1 comprises a block 9 of randomly selecting text parameters, the output of which is connected to the input of the text generation unit 10, and the output of which is connected to the input of the projective transformation unit 11, the output of which is connected to the input of the color variation block 12, output which is connected to the input of block 13 of the interference simulation, and

- a random text selection unit 9 is configured to randomly select text parameters and transfer them through the text generation unit 10 to the block 11 projective transformations, to the block 12 color variations, and to the block 13 to simulate interference,

- the text generation unit 10 is configured to form an image of the text and transfer it to

- block 11 of projective transformations, which is configured to convert the image of the text in accordance with the projective parameters and transfer it to

- block 12 chromaticity variations, which is adapted to convert the image of the text in accordance with the chromaticity parameters and transfer it to

- block 13 simulation of interference, which is made with the ability to convert the image of the text in accordance with the parameters of the noise, approximating the form of the generated text to the text that exists under actual conditions, as well as the ability to transfer the image of the text in the block seamless insert.

In the preferred embodiment of the system, the text parameters are selected from a set of parameters containing range parameters, viewing angle, perspective, color, sharpness, contrast, as well as interference parameters that approximate the appearance of the formed text to the text that exists in real conditions.

For a better understanding of the claimed utility model, the following is a detailed description with appropriate graphic materials.

FIG. 1. Scheme of the system for generating images containing text made according to the utility model.

FIG. 2. Scheme of the background image forming unit, made according to the utility model.

FIG. 3. Scheme of the formation of the image of the text, made according to the utility model.

FIG. 4. An example of a text image generated according to a utility model:

a) source text;

b) a generated image containing an image of the source text and a background image.

Items:

1 - block the formation of the image of the text;

2 - block forming the background image;

3 - block seamless insert;

4 - block variations of the final image;

5 - unit for learning recognition algorithms;

6 - base of background images;

7 - block selection of the background image;

8 is a block for selecting the scale and area of the background image;

9 - random selection of text parameters;

10 - text generation unit;

11 is a block of projective transformations;

12 - block variations of color;

13 is a noise simulation unit.

Let us consider in more detail the embodiment of the utility model shown in FIG. 1 - 4. Teaching image recognition algorithms requires a large number of text images (teaching examples), on which texts should be visible on a diverse background, while the background should correspond to the maximum of the actual background that will be recognized using a real optical recognition system. texts.

First, a random background image is separately formed using the background image forming unit 2 and the text image using the text image forming unit 1 (FIG. 1). Then, using the block 3 seamless insert, the text image is superimposed on the background image, at the same time the image defects at the borders are reduced, where the text image goes into the background image. After that, using block 4, variations of the final image modify the final image using image variation methods commonly used to train computer vision algorithms. After that, the modified image is transmitted to the external recognition algorithm learning unit 5, which is part of the external image recognition system.

While the images of the required texts can be synthesized programmatically, since their appearance is strongly formalized, the background image can differ in extreme diversity. At the same time, in many cases, the required background images in a variety can be collected from available sources (images on the Internet, publicly available data sets, other photos and video materials), so it is advisable to use such images from available sources as background images texts. In the simplest case, a text image is applied to a background image in a randomly selected location, with the result that the text image position is unrealistic, but this is a payment for the simplicity of the claimed system and does not interfere with the subsequent full use of the final image in external image recognition systems.

In block 2 of the formation of the background image perform the following operations (Fig. 2). Using the block 7 of the selection of the background image carry out a random sample of the background image of the pre-formed base 6 background images. Using the block 8 for selecting the scale and the background image area, a probabilistic zooming of the background image and a random sample of the background image area of a given size are performed as the final background image for the final image, after which the background image is transferred to the seamless insert block 3.

In block 1 of the formation of the image of the text perform the following operations (Fig. 3). For each generated image with the help of block 9 of random selection of text parameters, randomly select text parameters within the limits determined by a specific task. Then, using the text generation block 10, a text image is formed against a neutral background. Using a block of 11 projective transformations, random range parameters are selected, the angle of view is within the allowable range of the problem to be solved, and an appropriate perspective transformation is applied to the text image, resulting in the text image being formed at a given angle of view. Then, using block 12, chromaticity variations introduce random distortions of the color, sharpness, and contrast parameters of the text. After that, with the help of block 13, noise simulation is put on the image of the text of the noise, which approximates its appearance to that obtained during actual surveys.

Since the resulting text image generated by the functioning of the claimed utility model is highlighted in the background image by quality, white balance and other parameters, there is a risk that the external OCR system will use these features (which will not be in real terms). In order to reduce this effect, simultaneously with the generation of the required texts, by the same rules, false examples are formed that the external OCR system (including other format inscriptions and other object images), which are also applied to the background image, should not react. The obtained background images are used when teaching an external OCR system as false examples.

Since when converting an image of text in block 11 using a perspective transformation, distortions arise along the border of a text image, when superimposed on the background image, the border is clearly defined, which distinguishes the generated image from the real one. To combat this effect, use block 3 seamless insert, functioning as follows.

When forming the original image of the text, they simultaneously form a mask bounding it (mask value = 1 at points belonging to the text and 0 at points not belonging to the text). With further changes, the image simultaneously changes the mask, while where on the border the image is blurred between adjacent pixels, the mask is also washed out in the same way.

In order to make the borders less ideal-straight, within a given number a of pixels from the border (the specific value of a is several pixels and is a parameter of the method) at each point (pixel) the mask value is replaced with the minimum mask value in neighboring pixels located at a distance b from a given pixel, with the value of b ranging from 0 to a in each pixel being chosen randomly.

When forming the final image, the brightness I of each pixel is calculated using the formula I = m * I _f + (1-m) * I _b , where m is the mask value in the given pixel after the transformations described above, I _f and I _b , respectively, brightness text images after transformation and brightness of the background image.

As a result of the seamless insertion in the final image, there are no distortions that would be present if one image were simply replaced by another in a certain area, besides the border is blurred on the final image, as it happens on real images.

The claimed utility model solves the problem of the laboriousness of preparing initial data for the process of learning text recognition algorithms. In the claimed utility model, the characteristic time for generating one image is less than 1 second, which is thousands of times less than the preparation time for manually-labeled images, especially considering the time required to organize and conduct the filming of texts with different parameters.

Although the above described embodiment of the utility model was set forth to illustrate the claimed utility model, it is clear to specialists that various modifications, additions and substitutions are possible without departing from the scope and meaning of the claimed utility model disclosed in the attached utility model formula.

Claims (15)

1. The system for generating images containing text containing a text imaging unit 1 and a background image forming unit 2, the outputs of which are connected to the input of the seamless insert unit 3, the input of which is connected to the input of the final variation image 4, and - block 2 forming the background image made with the possibility of forming the image of the background and transfer it to the block seamless insert, - block 1 of the imaging text is made with the possibility of forming an image of the text and transfer it to - block 3 seamless insert, which is made with the possibility of forming the final image by overlaying the image of text on the background image, while reducing image defects on the border of the text image and the background image, as well as the ability to transfer the final image to - block 4 variations of the final image, which is made with the possibility of modifying the final image using image variation methods used to train computer vision algorithms, as well as the ability to transfer the modified final image to the external block 5 learning recognition algorithms, which is part of the external image recognition system . 2. The system of claim. 1, characterized in that the background image forming unit 2 comprises a background image selection unit 7, the input of which is connected to a preformed background image base 6, and the output of which is connected to the input of the scale selection unit 8 and the background image area, where - the block of the choice of 7 background images made with the possibility of a random selection of the background image of the pre-formed base 6 background images and transfer it - in block 8 of the choice of scale and the background image area, which is made with the possibility of a probable change of the background image scale and a random selection of the background image area of a given size as the final background image for the final image, as well as the ability to transfer the background image into the seamless insert unit 3. 3. The system of claim 1, wherein the text imaging unit 1 comprises a block 9 of randomly selecting text parameters, the output of which is connected to the input of the text generation unit 10, and the output of which is connected to the input of the projective transformation unit 11, the output of which is connected to the input of the block 12 chromaticity variations, the output of which is connected to the input of the block 13 of the interference simulation, and - a random text selection unit 9 is configured to randomly select text parameters and transfer them through the text generation unit 10 to the block 11 projective transformations, to the block 12 color variations, and to the block 13 to simulate interference, - the text generation unit 10 is configured to form an image of the text and transfer it to - block 11 of projective transformations, which is configured to convert the image of the text in accordance with the projective parameters and transfer it to - block 12 chromaticity variations, which is adapted to convert the image of the text in accordance with the chromaticity parameters and transfer it to - block 13 simulation of interference, which is made with the ability to convert the image of the text in accordance with the parameters of the noise, approximating the form of the generated text to the text that exists under actual conditions, as well as the ability to transfer the image of the text in the block seamless insert. 4. The system of claim 3, wherein the text parameters are selected from a set of parameters containing range parameters, viewing angle, perspective, color, sharpness, contrast, as well as interference parameters that approximate the appearance of the formed text to the text that exists in real conditions .

Images