Device and method for recognition of persian number plates in motion

Abstract

The present invention relates to a method for the recognition of Persian number plates in motion, comprising: taking a video sequence in real time of either a moving or parked vehicle, using a camera incorporated into a device in a vehicle in motion; multiple detection processing of all the number plates of all the cars that appear in the scene; selecting the image to be manually processed, the image containing the number plate to be processed and analysed; segmenting the selected image; applying optical character recognition to extract the segment of each Persian character; post-processing to detect errors and detect special characters such as those used by taxis and vehicles for the disabled; and verifying the extracted data in a local or remote repository, involving querying an external database. The invention further relates to a device for the recognition of Persian number plates in motion.

Description

[0001] DESCRIPTION

[0002] Device and method for recognizing moving Persian license plates

[0003] Technical field

[0004] The present invention relates to a new type of device designed to be installed in a vehicle and to be able to capture the data in motion necessary to carry out a method of recognizing Persian license plates by means of image processing.

[0005] The invention is framed within the different types of equipment and methodologies for license plate recognition, the present invention being oriented to the particular recognition of license plates with Persian characters.

[0006] State of the art

[0007] As is well known, current license plate image recognition systems run on devices such as cameras, data transmission equipment, storage systems, or others, based on small, decentralized pieces of software capable of recording moving images and sending them, via the communications network, to a central recognition system. These solutions, such as those disclosed in document US2023004747 or document KR102480582, are distributed systems in which multiple devices are interconnected by their own communications network or the Internet, capable of cooperating in solving the problem by exchanging messages, while maintaining their own memory and processor autonomy.

[0008] In this regard, the applicant is not aware of any distributed system that allows for the processing and identification of license plates on the move, specifically those based on Persian characters.

[0009] For moving scenarios without an Internet connection or a standard communications network and requiring crystal-clear images, this type of distributed system is not viable, as it requires a high-performance communications network for image transmission and processing, and cannot perform license plate identification in real time (at the same time as the photo is taken).

[0010] A typical example scenario would be a user taking a photo with a mobile phone or mobile radar on a dirt road with dust and adverse lighting conditions, and who needs to identify a license plate with Persian characters at the time of taking the photo. As previously mentioned, this is not feasible with distributed systems known in the art.

[0011] Therefore, if under these conditions, identification of license plates based on Persian characters is desired with the mobile device with which the photo is taken, this is not feasible with current mobile systems for the following reasons:

[0012] (i) would require high energy consumption that would be unfeasible with the current battery capacity of these devices. Furthermore, energy consumption increases exponentially in adverse lighting conditions such as low light, smudges on the license plate, rain, and other elements that make it difficult to illuminate the license plate;

[0013] (i) would have a very low level of efficiency since real-time license plate identification requires a higher computing capacity than any current mobile device allows. Their efficiency would be far below that of Distributed Systems that have time to process and identify license plates.

[0014] (iii) There are no mobile systems capable of identifying license plates based on Persian characters in real time.

[0015] Existing solutions for mobile devices using SSD (Single Shoot Device) models are capable of identifying objects in a single image processing, which, in principle, makes them feasible for real-time recognition. However, the subsequent process involving license plate detection, correcting perspective deviations by projecting the real plane onto the camera plane, segmentation and extraction of Persian characters, filtering, cleaning, and OCR (Optical Character Recognition) does not allow for the effective detection of license plates with Persian characters using the previous technique.

[0016] On the other hand, both the distributed and mobile systems analyzed are primarily based on the Latin alphabet, the most widely used writing system in the world. As mentioned previously, there are currently no solutions for Persian characters. It's worth clarifying that a model trained with images of Latin characters or Chinese characters with highly defined lines will never work to classify images based on Persian characters. This is because these systems don't use a database with a character dictionary, but rather recognition models that have been trained with the images they must then classify.The presence of smudges on license plates, dirt, rain, noisy elements such as the screws used to secure the license plate, and even superimposed elements that are intentionally used to circumvent identification systems, in addition to adding complexity and energy consumption, make the deployment of a system that simply swaps Latin or Chinese characters for characters from another language unviable.

[0017] More specifically, regarding Persian characters, their characteristics add greater complexity to recognition. For example, the Persian numeral zero (a small diamond) in a photograph is very similar to the back of a screw. Furthermore, Persian license plates also include characters that are very difficult to identify due to the presence of umlauts in some cases, and the features of the writing itself.

[0018] The present invention is capable of processing moving images at the speed required for real-time license plate recognition applications, recognizing Persian license plates in low-light environments on both mobile and fixed devices. The developed method and models are optimized so that the entire identification and recognition process is carried out, indistinctly, with an architecture of minimum requirements that is easily achievable and that can be covered with relatively low-cost products such as devices known in the market, such as a Raspberry Pi 4, or mobile devices such as a Blackview BV9700 mobile phone. In addition, the present technology takes advantage of hardware accelerators such as GPUs or processors optimized for AI (Artificial Intelligence) that are available in the device.

[0019] This invention, unlike any other known technology in this industrial sector, provides the first market-leading approach to license plate detection based on Persian characters that operates autonomously on mobile devices without the need to be connected to a communications network. Another key feature that distinguishes this invention from other existing technologies on the market is the device's ability to detect and recognize license plates completely autonomously, disconnected from the Internet, with optimal power consumption, and with an efficiency, as explained in this document, between 90 and 95%.This is possible thanks to the invention of a new method tailored to the recognition, segmentation, and optical recognition of Persian characters. This method relies on various optimizations in the artificial intelligence models it uses to take advantage of the device's hardware accelerators, namely its processor and graphics unit.

[0020] In this regard, the applicant is not aware of a solution as effective to this problem of identifying Persian license plates as the one described and claimed below.

[0021] Explanation of the invention

[0022] The present invention, as previously mentioned, relates to a new type of device designed to capture the moving data necessary to carry out a method for recognizing Persian license plates by means of image processing. That is, one object of the invention is based on the definition of the device, and another object of the invention is based on the definition of the aforementioned method for recognizing Persian license plates.

[0023] The device object of the present invention has the particularity compared to any other known in this industrial sector of being constituted by a series of elements that allow the previously indicated method to be developed, and this device comprises: a screen, which is a touch display system; a central processing unit NPU, which is a processor for mobile devices optimized to execute Artificial Intelligence models based on Tensorflow-Lite type components, with hardware accelerators based on GPUs (Graphics Processing Units) and based on a 64-bit ARM type instruction set architecture; a graphics processor, which is a GPU processing unit integrated into the central processor such that it manages the graphic operation of the mobile device; a memory unit, preferably RAM (Random Access Memory) with at least 3GB capacity.a persistent data storage unit, a Nano-SIM or Nano-SIM / microSD-based system for storing video and image sequences; a communication module, which may be any of the remote connections; an information exchange camera, which is a unit with a built-in or connected camera; and an autonomous power system, which is a battery, preferably a lithium-ion polymer battery.

[0024] The mobile device enables the development of a novel method for multiple character detection and recognition, as well as the automatic extraction of all Persian characters (digits and numbers). Tests have determined an accuracy rate of over 90%. This figure is the result of a series of tests using a random sample of 20% of the total images used, out of approximately 50,000. The images used include vehicles in various lighting conditions (night and day scenes), rain, dust, partial license plate visibility, multiple vehicles, and others.

[0025] The method is specifically designed for license plates with Persian characters, which is unique compared to other types of license plates, in addition to Persian characters, that they include special symbols, such as those used to identify people with reduced mobility, special vehicles, or taxis. As previously indicated, there are known technologies that can detect license plates with other types of characters, for example, Chinese, but these technologies do not allow the detection of Persian characters. Additionally, the invention also allows for handling different types of colors, which are used for protocol vehicles, both public and private.

[0026] The proposed method enables real-time vehicle identification, immediately with a latency of around 100 ns, on the autonomous device, at multiple angles, in adverse lighting conditions such as low light, dust and rain, foreign objects on the license plate, and fog; and the recording of traffic data, for example, owner or vehicle data. In the case of scene conditions, no additional adjustment is required, regardless of these conditions. The method allows the device to be used for license plate recognition and to integrate Artificial Intelligence (AI) mechanisms into a mobile device. Without this method, it would not be viable due to the high execution cost and short-term battery drain, as previously mentioned, which is a problem with current proposals.The method provides the device with identification and recognition functionality, the necessary adjustments to adapt to the operating height in correspondence with the size of a person, such as image correction based on the angles at which the image has been taken, and multiple inclinations without this being an aspect that could affect it.

[0027] The method, which is performed autonomously on the device, consists of an embedded neural network system optimized to take advantage of the hardware accelerators built into a Central Processing Unit (NPU), capable of performing multiple detections of moving objects with a latency of less than 100 ms. The built-in inference module is capable of detecting small objects in low-light environments, with different perspectives and high levels of occlusion.

[0028] In a schematic and summary form, the method consists of the following stages:

[0029] Stage 1: In this stage, the video sequence is taken in real time, whether it is a moving or parked / stopped vehicle.

[0030] The method begins by capturing a video sequence from the vehicle's built-in camera while it is moving, focusing on the targets, in this case, vehicles that are either stationary or moving.

[0031] Stage 2: The multiple detection process is performed for all license plates of all cars appearing in the scene of each of the video frames resulting from the previous stage. In this stage, the level of certainty of each detection is indicated, with green and red colors indicating whether or not a predetermined certainty threshold (90% or higher) is exceeded, respectively.

[0032] That is, there is multiple detection of license plates in each frame, in which the rectangle of each detected license plate is displayed with a % accuracy indicator. The neural network model used is optimized; inference is made using the input video frames (TFLite with GPU delegates); and models such as Mobilenet_SSD_COCO_v2 (TF Mobile and TFLite versions).

[0033] The entire workflow is performed on the mobile device. The steps are as follows: a) Quantization of the neural network model (approximately 15 MB); b) Calculation of the camera plane angles (XYZ) from intrinsic properties (focal length, optical center); c) During execution using hardware accelerators (latency less than 100 ms).

[0034] Step 3: The user must select the image to be processed with their finger. This image will contain the license plate they wish to process and analyze.

[0035] Step 4: The selected license plate is cropped, and a cropping process, Cartesian or homographic transformation, is performed to correct distortions due to the perspective in which the device is oriented with respect to the scene.

[0036] Cropping and saving to a temporary folder includes: i) Correcting skew / deformation.

[0037] i) Filtering to eliminate noise (shadows, dirt, auxiliary elements to fix the license plate, etc.) iii) Binarization of the image (contrast increase) iv) Detection of character segments (8 in total)

[0038] This process begins by obtaining a matrix of values ​​for the intrinsic properties of the device's camera. The Euler angles are calculated using the intrinsic parameters (the camera is calibrated using checkerboard patterns to obtain the intrinsic matrix K. Using the homography matrix H, we can calculate the angles known in English as pitch, roll, and yaw). The cropping, transformation, and correction process is based on the following steps:

[0039] Adding canvas

[0040] Converting to grayscale

[0041] Filter 2D, Gaussian and bilateral filtering.

[0042] Image binarization with OTSU dynamic filter

[0043] Detect maximum contour (convex hull)

[0044] Detect corners from minAreaRect of Max Contour

[0045] Calculate destination points

[0046] Unboxing image using homography matrix

[0047] Check the aspect ratio (width, height) and crop the image by adjusting the aspect ratio to 1:2

[0048] Once the cropping has been corrected, the character segmentation process begins, which involves determining the rectangle that best fits each character. Once each segment (a Persian character framed within a rectangle) is obtained, it is cropped, and dynamic and morphological filters are applied, correcting and cleaning the character, removing noise, shadows, distortions, etc.

[0049] Step 5: Optical Character Recognition (OCR) is applied to extract the character in Char format (1 byte length) from the segment of each Persian character using the OCR software Tesseract.

[0050] This Optical Recognition process is performed only once, with automatic correction of the central character and the set of digits (e.g., similar characters)

[0051] Stage 6: Post-processing is performed to detect errors in the previous stage, and to detect special characters such as those used by taxis and the disabled.

[0052] During this stage, the character histograms are analyzed to identify which ones constitute special characters. By analyzing the histogram patterns, we compare them with those of each of the possible characters (taxi, disabled person).

[0053] Step 7: Verification of the extracted data in a local or remote warehouse.

[0054] Query an external database to extract data from license plates of interest, for example, owner information, address, whether or not they owe fines, etc.

[0055] Optionally, step 8: Determining quality metrics.

[0056] Developing all the steps in a compact and concise manner, first, the entire video sequence is captured with the device image. During detection, multiple video frames are processed, and the license plates detected in the scene are framed with a certainty indicator. The user then selects the image containing the license plate to be analyzed. From the selection, the area of ​​interest is processed and cropped, and the characters are segmented. The OCR (Optical Character Recognition) component is equipped with tools to read the image segments of each character and generate the corresponding text string. During post-processing, special characters, such as those incorporated in taxis, cars for the disabled, etc., are verified. This is a customizable process where country-specific rules governing the arrangement of characters within the license plate can be added.Finally, the corresponding verification is performed by consulting external data sources. As previously indicated, recognition efficiency is around 90% successful for all characters in the worst-case scenarios. Manual correction of the read license plate is also possible. The final stage, which is optional, consists of determining the process quality parameters by analyzing the intrinsic properties of the device's built-in camera.

[0057] The device's method is performed in real time, capable of processing multiple video frames at a range of 40 to 60 fps. It effectively corrects distortions that arise from the license plate due to the difference in the scene and camera planes, resulting from the camera's rotation relative to its Euler angles. One of the most relevant aspects to highlight is that the entire process has been optimized to improve the device's battery performance, drastically reducing its consumption.

[0058] The device applies multiple filters (morphological, Gaussian, etc.) to the image to eliminate elements of the license plate related to the presence of dirt, external agents, or others.

[0059] A histogram analysis is performed to equalize the brightness and contrast of the detected image. The area of ​​interest is then analyzed, and characters are extracted, which are then processed by OCR and refinement according to specific rules of the Persian alphabet, which involve combining numbers or characters to denote codes such as locations and others.

[0060] In this regard, and as seen in Figs. 6A-6D, a sample of 34 images for license plate recognition is provided, showing the results of Detection, Homographic Transformation, and Segmentation (although tests have been performed on approximately 50,000 images). The 90% effectiveness is relative to the entire process. This means that in 9 out of 10 frames of the video sequences taken, the method we are recording concludes successfully. This means several things: first, that the detection works correctly, that cropping can then be performed, that image distortion can be corrected (homographic transformation), and then that the OCR performs the recognition; and second, that the user can decide which license plate to recognize, manually, by selecting the appropriate rectangle.

[0061] The device is designed to optionally connect to external points to extract owner information and perform specific actions related to fines, reports, etc. The interface is simple, allowing the operator to select the license plates of interest and make corrections manually. The developed system is designed to store the necessary information, such as failure rates, images, and results of method executions, to continuously improve the neural network models used and perform transfer learning.

[0062] The Applicant is not aware of a customized recognition device or method for moving Persian license plates as effective as the one previously described. Furthermore, it should be noted that, throughout the description and claims, the term "comprises" and its variants are not intended to exclude other technical features or additional elements.

[0063] Brief description of the figures

[0064] In order to complete the description and to help better understand the characteristics of the invention, a set of figures and drawings is presented in which the following is represented for illustrative and non-limiting purposes:

[0065] Figure 1 shows schematically the structure of the device object of the present invention.

[0066] Figure 2 shows a schematic flowchart showing the steps of the method object of the present invention.

[0067] Figure 3 visually shows a sequence of the license plate deformation correction process, with tilt angles greater than 45 degrees, which allows the orientation of the license plate to be rectified before starting the segmentation and character extraction process.

[0068] Figure 4 visually shows a sequence of the results of the license plate deformation correction process and the subsequent character segmentation, with the rectangles (known in English as Bounding Boxes) corresponding to each detected character.

[0069] Figure 5 visually shows a sequence where the process of transformation and filtering of the license plate is applied from its binarization: transformation of the original image in colors to grayscale, obtaining the inverted binary image (only in black and white), detection of edges and contours, painting of contours, conversion to the black and white image.

[0070] Figures 6A-6D show several examples of license plate recognition with homographic transformations and segmentations, and illustrate some errors in the recognition that can be manually corrected. Detailed explanation of an embodiment of the invention

[0071] Next, a possible embodiment is briefly described, firstly, of the device for recognising moving Persian number plates which, as can be seen in Fig. 1, comprises: a screen (1), which is a touch display system, of at least 6.53 in, IPS, 720 x 1600 pixels, 24 bits; a Central Processing Unit NPU (2), which is a processor for mobile devices optimized to run Artificial Intelligence models based on Tensorflow-Lite, with hardware accelerators based on GPUs (Graphics Processing Units) and based on 64-bit ARM architecture; a Graphics processor (3), which is a GPU processing unit integrated into the central processor so that it manages the graphic operation of the mobile device; a memory unit (4), RAM memory of at least 3GB capacity; a persistent storage unit (5), a system based on Nano-SIM or Nano-SIM / microSD for storing video and image sequences; a communication module (6), any of the Wi-Fi connections a, b, g, n 5 GHz, Dual Band, Wi-Fi Hotspot, Wi-Fi Direct, Bluetooth 4.2, micro USB 2.0 an Information Exchange Unit (7): built-in or connected camera, at least 3264 x 2448 pixels, 1920 x 1080 pixels, 30 fps; an autonomous power system (8) which is a battery of at least 4780 mAh, lithium-ion polymer.

[0072] The Persian license plate recognition method, as shown in Fig. 2, comprises the following steps:

[0073] (ET1) a real-time video sequence shot, whether of a moving or parked / stopped vehicle, with a camera incorporated into a device in a vehicle;

[0074] (ET2) a multiple detection processing of all the license plates of all the cars that appear in the scene of each of the video frames resulting from the previous stage; where the rectangle of each of the detected ones is shown with a % certainty indicator, and where there is an optimization of the neural network model used; having an inference using the input video frames (TFLite with GPU delegates); and using models such as Mobilenet_SSD_COCO_v2 (TF Mobile and TFLite versions). The entire flow is performed on the mobile device itself and includes the following steps: a) Quantization of the neural network model b) Calculation of the camera plane angles (XYZ) from the intrinsic properties of focal length and optical center. c) During execution, hardware accelerators with latency less than 100 ms are used

[0075] In one embodiment of the invention, this is done on a Raspberry type device, although it can also be executed on a Blackview BV9600, BV9700 Pro device, including Neuro Pilot AI with Helio P60, P70.

[0076] (ET3) a selection of the image to be processed manually, specifically by fingerprint, where the image contains the license plate you wish to process and analyze.

[0077] (ET4) a segmentation of the selected image, where the selected license plate is cropped, followed by a Cartesian or holographic transformation, to correct distortions due to the perspective in which the device is oriented with respect to the scene; cropping and saving in a temporary folder includes: i) Correction of tilt / deformation.

[0078] i) Filtering to eliminate noise (shadows, dirt, auxiliary elements to fix the license plate, etc.) iii) Binarization of the image (contrast increase) iv) Detection of character segments (8 in total)

[0079] Where we start by obtaining a matrix of values ​​of the intrinsic properties of the device's camera, the Euler angles are calculated, using the intrinsic parameters (the camera is calibrated using chess patterns to obtain the intrinsic K matrix. Using the homography matrix H we can calculate the angles known in English as Pitch, Roll and Yaw).

[0080] And where the cropping, transformation and correction is based on the following steps:

[0081] Adding canvas

[0082] Converting to grayscale

[0083] Filter 2D, Gaussian and bilateral filtering.

[0084] Image binarization with OTSU dynamic filter

[0085] Detect maximum contour (convex hull)

[0086] Detect corners from minAreaRect of Max Contour

[0087] Calculate destination points

[0088] Uncrop image using homography matrix Check aspect ratio (width, height) and crop image by adjusting aspect ratio to 1:2

[0089] Once the cropping has been corrected, the character segmentation process begins, which involves determining the rectangle that best fits each character. Once each segment (a Persian character framed within a rectangle) is obtained, it is cropped, and dynamic and morphological filters are applied, correcting and cleaning the character, removing noise, shadows, distortions, etc.

[0090] (ET5) application of an optical character recognition (OCR) to extract from the segment of each Persian character, using the OCR software Tesseract, the character in Char format (length of 1 byte), where this process is performed only once, with automatic correction of the central character and the set of digits (e.g., similar characters j^)

[0091] (ET6) a post-processing to detect errors in the previous stage, and detect special characters such as those used by taxis and disabled people, where in this stage the histograms of the characters are analyzed, to detect which ones constitute special characters, and analyzing the patterns of the histogram, it is compared with those of each of the possible characters (taxi, disabled person).

[0092] (ET7) a verification of the extracted data in a local or remote warehouse, where an external database is consulted to extract data from the license plates of interest, for example, the owner's data, address, whether or not they owe fines, etc.

[0093] (ET8) a determination of the quality metrics, which is an optional step, which in one embodiment of the invention is carried out with a commercial type Sony IMX219 8-megapixel camera compatible with the Raspberry P¡-4 device and with a 4608x3456 pixel, 30fps camera of the Blackview BV9700 device, to determine the quality metrics, the Euler angles are calculated using the intrinsic properties of the camera, where the camera is calibrated with checkerboard patterns to determine the intrinsic matrix K, and through this matrix K and the homographic matrix H the Pitch, Roll and Yaw angles are calculated.

[0094] Developing all the stages in a compact and summarized way, first, (ET1) the complete video sequence is captured with the image of the device; (ET2) during detection, multiple video frames are processed and the license plates detected in the scene are framed with a certainty indicator; (ET3) then, the user selects the image that contains the license plate object of analysis; (ET4) from the selection, the area of ​​interest is processed and cropped and the characters are segmented; (ET5) the OCR (Optical Character Recognition) component is equipped with tools to read the image segments of each character and generate the corresponding text string; (ET6) during post-processing, special characters are verified, such as those incorporated in taxis and cars for the physically disabled;(ET7) Finally, the corresponding verification is performed by consulting external data sources. As previously indicated, the recognition efficiency is around 90% successful for all characters in the worst-case scenarios; and (ET8) where the determination of the process quality parameters is also carried out by analyzing the intrinsic properties of the camera incorporated in the device.

[0095] It is worth noting that Fig.3 visually shows a sequence of the license plate deformation correction process, with inclination angles greater than 45 degrees, which allows the orientation of the license plate to be rectified before starting the segmentation and character extraction process, where it starts with the original image (3.1.), moving on to a filtered image (3.2.), to subsequently apply an OTSU type threshold processing (3.3.), determine the polygon outline (3.4.), determine the corner points (3.5), obtain a cropped image (3.6.) and obtain the crop (3.7.)

[0096] As for Fig.4, it visually shows a sequence of the results of the license plate deformation correction process and the subsequent segmentation of the characters, with the rectangles (known in English as Bounding Boxes) corresponding to each detected character, where it starts from a starting image (4.1.), moving on to a corrected image (4.2) and an image with the segmentation (4.3.)

[0097] Continuing with Fig. 5, this figure visually shows a sequence where the process of transformation and filtering of the license plate is applied from its binarization, from a starting image (5.1.) to the transformed image (5.2.), where there is a transformation of the original image in colors to grayscale, obtaining the inverted binary image (only in black and white), detection of edges and contours, painting of contours, conversion to the black and white image.

[0098] Finally, Figures 6A to 6D show an example of 30 license plate recognition samples, starting from a detected image (6.1) from the beginning, obtaining images of the homographic transformations (6.2) and obtaining images of the segmentations (6.3).

Claims

CLAIMS 1.- Method for recognizing Persian license plates in motion, comprising: (ET1) a real-time video sequence shot, whether of a moving or stationary vehicle, with a camera incorporated in a device in a moving vehicle; (ET2) a multiple detection processing of all the license plates of all the cars that appear in the scene of each of the video frames resulting from the previous stage; where there is an optimization of a Mobilenet_SSD_COCO_v2 type neural network model; where the process is carried out on the mobile device itself and includes the following steps: a quantization of the neural network model; a calculation of the camera plane angles (XYZ) from the intrinsic properties of focal length and optical center; the use of hardware accelerators with latency less than 100 ms; (ET3) a selection of the image to be processed manually, specifically by fingerprint, where the image contains the license plate you wish to process and analyze. (ET4) a segmentation of the selected image, where the selected license plate is cropped, a Cartesian or homographic transformation is performed to correct distortions due to the perspective in which the device is oriented with respect to the scene; and a segmentation of the characters, which consists of determining the rectangle that best fits each Persian character framed in a rectangle, which is cropped and dynamic and morphological filters are applied, which correct and clean the character, (ET5) an application of optical character recognition (OCR) to extract from the segment of each Persian character, using the OCR software Tesseract, where this process is performed only once, with automatic correction of the central character and the digit set; (ET6) a post-processing to detect errors in the previous stage, and detect special characters such as those used by taxis and the disabled, where in this stage the histograms of the characters are analyzed, to detect which ones constitute special characters, and analyzing the patterns of the histogram, they are compared with those of each of the possible characters; and (ET7) a verification of the extracted data in a local or remote warehouse, where an external database is consulted 2.- A method, according to claim 1, characterized in that it comprises a step final (ET8) with a determination of quality metrics by analyzing the intrinsic properties of the camera built into the device. 3.- A method according to claim 1, wherein in the segmentation stage (ET4), the cropping of the image and segmentation in a temporary folder comprises: i) Correction of the inclination / deformation. i) Filtering to remove noise such as shadows, dirt, auxiliary elements to fix the license plate or others; iii) Binarization of the image based on contrast enhancement; iv) Detection of Persian character segments, which are 8 in total. 4.- A method, according to claim 3, where after the cropping a transformation is made where a matrix of values ​​of the intrinsic properties of the device's camera is obtained, where the Euler angles are calculated using intrinsic parameters based on the camera being calibrated using chess patterns to obtain the intrinsic K matrix, and using the homography matrix H the angles known in English as Pitch, Roll and Yaw are calculated. 5.- A method, according to claim 4, where after the transformation a correction is made based on the following steps: Adding canvas Conversion to grayscale Filter 2D, Gaussian and bilateral filtering. Image binarization with OTSU dynamic filter Detect maximum contour (convex hull) Detect corners from minAreaRect of Max Contour Calculate destination points Unboxing image using homography matrix Check the aspect ratio (width, height) and crop the image by adjusting the aspect ratio to 1:2 6.- Device for moving Persian license plates according to a method according to any of claims 1 to 5, characterized in that it comprises: a screen (1), which is a touch display system; a Central Processing Unit NPU (2), which is a processor where Artificial Intelligence models based on Tensorflow-Lite are executed, with hardware accelerators. based on GPUs (Graphics Processing Units) and based on 64-bit ARM architecture; a Graphics processor (3), which is a GPU processing unit integrated into the central processor where the graphic operation of the mobile device is managed; a memory unit (4); a persistent storage unit (5) for storing video and image sequences; a communication module (6); an Information Exchange unit (7), with a built-in or connected camera; and an autonomous power system (8). 7.- A device according to claim 6, wherein the memory unit (4) is a RAM memory of at least 3GB capacity; 8.- A device according to claim 6, wherein the persistent storage unit (5) is a Nano-SIM or Nano-SIM / microSD based system. 9.- A device according to claim 6, wherein the communication module (6) is a connection module with Wi-Fi a, b, g, n 5 GHz, Dual Band, Wi-Fi Hotspot, Wi-Fi Direct, Bluetooth 4.2, micro USB 2.0 10.- A device according to claim 6, wherein the autonomous power system (8) is a battery of at least 4780 mAh. 11.- A device according to claim 10, wherein the battery is a lithium-ion polymer.

Images