TWM618756U - Image recognition system - Google Patents

Abstract

The present utility proposes a novel image recognition system, comprising the following modules. (i) A control module that controls the operation of the object character recognition system. (ii) An image preprocessing module, coupled with the control module, that adjusts the quality of image. The image preprocessing module further contains a positioning unit that allocates the graphic coordinate within the image. (iii) An identification module, coupled with the image preprocessing module, that adjusts the graphic area to be processed in the image via the graphic coordinate and outputs recognition results according to the graphic area.

Description

Image recognition system

This creation relates to an image recognition system. More specifically, it is a recognition system that compares the image data to be recognized by the recognition module, and then translates the image text or symbol corresponding to the image data.

In recent years, as hardware computing capabilities have become more sophisticated, and scanning instruments and handheld devices have become popular, image recognition technology has been widely used in the market. Among them, the so-called image recognition technology is a technology for recognizing image data. A common application is to use a scanning instrument to input text or images in a specific format into the computer through optical scanning, and after a specific code is recognized, it replaces the user’s manual input, thereby shortening Working time, which is of great help to the processing of huge printed materials. In addition, compared to the aforementioned optical scanning, there are also handwriting recognition systems for handwritten fonts or symbols currently on the market. The symbols or texts can be recognized by text input such as handwriting pads, touch pads, and touch screens. .

Optical scanning technology was first used to process a large number of newspapers and magazines, documents, receipts and reports, etc. For organizations with a large number of paper documents, using text recognition technology to digitize documents can not only reduce storage space, but also set electronic tags for various documents to facilitate classification management. Therefore, the development of printed text recognition is relatively early. In the case of printed Chinese characters, the research on input using optical scanning technology was first published by two IBM researchers, Casey and Nagy, in an IEEE journal in 1966 on Chinese character recognition. In this paper, by splitting the fonts of printed Chinese characters into the template of the bipolar matrix to build a database, and then using the content of the database to perform template matching on the objects to be recognized, the recognition is finally achieved Research results of 1,000 printed Chinese characters.

In addition, for the method of recognizing text on handwritten documents, the previous related technology in Taiwan During the operation, the Industrial Technology Research Institute also disclosed a method of identifying Chinese and English forms through OCR in Taiwan Patent Publication No. TW 294804. The patent includes the use of the form learning module to learn the position coordinates of the fields, characters, grid lines, etc. of the form to be recognized, so as to enhance the accuracy of positioning the field where the character to be extracted is located, and to correct the form scanning The tilt and offset of the time; after extracting the characters in the predetermined field, a single comparison algorithm is used according to the needs of the application (such as the detailed description of the preferred embodiment in the '804 case, C. data extraction, 3. Character extraction, (ii) handwritten character extraction as described in the subsection), or comparison of several comparison algorithms with the characters stored in the database, and the confidence score after the comparison determines whether it needs to be corrected manually .

In view of the above, for optical scanning, text or images are included, and the most important factor that affects the recognition rate of subsequent text recognition is the extraction of its features and the comparison with the features. In recent years, with the rise of machine learning and the support of hardware equipment, Convolutional Neural Networks (CNN) has become the mainstream of feature extraction in image recognition. Common convolutional neural network structures are mostly composed of one or more convolutional layers and a fully connected layer at the top. They also often include associated weights and pooling layers. By sliding on the input pixel matrix, In the process of each convolution kernel corresponding to different features, a convolution operation (Convolution) is performed between the convolution kernel and the input pixel matrix, that is, the product of the pixel matrix and the corresponding elements of the convolution kernel is performed and summed, so as The information of a specific area on the pixel matrix is projected into the feature matrix, and features such as straight lines, horizontal lines, diagonal lines, circles, etc. are extracted.

However, although the above-mentioned text recognition technology has made great progress in recent years, it can achieve a very high recognition rate for images or text printed in a fixed and neat format, but for relatively scribble or large changes, especially handwritten text Because each person’s writing habits and handwriting are not the same, the recognition rate is low. It can only be used for entertainment and fun, or the recognition result can be corrected manually after recognition, but it cannot be officially used. This is especially true when the recognized text needs to be applied to formal documents, such as bank checks, contract documents, and legal documents. In addition, in the previous technology, there is also a lack of applications to convert the recognized text into another format after the interpretation of the text. Chinese characters, such as "five thousand zero one rounded", "five thousand zero one rounded", "five thousand zero one yuan", "five thousand and one yuan rounded" and other characters are converted into Arabic numerals corresponding to the same meaning " "5001", resulting in business applications, still need to spend a lot of human resources to review more conventional documents. Therefore, at the current point in time, there is still room for further improvement in the technology of image and text recognition.

In view of the shortcomings of the aforementioned conventional technology, this creation proposes an image recognition system. The system architecture includes: a control module, which provides system operation and management; an input module, which inputs image data; an image pre-processing module, which is coupled Connect the input module to adjust the data quality of the image data. The image pre-processing module includes a positioning unit to locate the image and text coordinates of the image data; the recognition module is coupled to the image pre-processing module to decompose the image data into multiple Decompose the data, and output the identification data to the control module according to the aforementioned graphic coordinates. Among them, the image data referred to in this creation can include words or symbols printed or handwritten on various documents and objects, such as "Five thousand and one round"; in addition, the identification data referred to in this creation includes After the above-mentioned characters or symbols are recognized by the image recognition system, the output formats, such as txt., doc., docx., xlsx, correspond to printed or handwritten characters or symbols, such as "Five thousand and one rounded", or Arabic The number "5001". Among them, in an embodiment of this creation, the length of the Arabic numerals can range from 0 to 20 digits.

According to the content of this creation, the above-mentioned image pre-processing module further includes a cutting unit, which cuts the image data into the image and text area to be recognized according to the image and text coordinates, and transmits the recognition result of the image and text area back to the recognition module to improve The graphic area and the accuracy of the graphic coordinates located by the positioning unit.

According to the content of this creation, the above-mentioned image pre-processing module further includes a correction unit, which determines the image angle or image color of the image data according to the image and text coordinates, and transmits the recognition result of the image angle or image color back Recognition module to improve recognition accuracy.

According to the content of this creation, the above-mentioned image pre-processing module further includes a classification unit, which determines the image type of the image data according to the image and text coordinates, and transmits the recognition result of the image type back to the recognition module to improve the image type The accuracy of recognition.

According to the content of this creation, the recognition module is based on the Encoder-Decoder architecture and includes an encoding unit, which converts image data into vector form encoded data through a feature extraction matrix; the decoding unit is based on the encoded data , Calculate the feature vector of the encoded data and output it as the decoded data.

According to an embodiment of the present invention, the image recognition system can optionally include a data storage module, coupled to the control module, and store the decoded data and/or the recognition data as training data according to the needs of the application. Among them, the decoded data includes graphic coordinates, graphic regions, graphic angles, graphic colors, and graphic types.

In an embodiment of this creation, the above positioning unit takes an application scenario as an example. It can have a step of automatic positioning and recognizing the target position (Automatic Positioning). The input image data (such as a check) can be imported first according to the needs of the application. Cut out a part through the cutting unit, such as the upper half or the lower half. The positioning unit outputs a graphic coordinate according to the area cut by the cutting unit, such as ( x, y ), and based on the graphic coordinate, the range that needs to be recognized is cut up, down, left, and right according to a preset value. For example, the horizontal axis is from x- △x to x + △x , and the vertical axis is from y- △y to y + △y . Recognizing this range can indicate a graph such as "5001" or "5001". The scope of the text is then used for the subsequent identification process based on the scope cut by the cutting unit.

In addition, in order to improve the shortcomings of the conventional technology, this creation also proposes an image recognition method. The method steps include inputting the image data and training data to be recognized through the input module; inputting the image data and training data into the recognition module , And by comparing the image data and training data to adjust the vector weights in the coding unit and the decoding unit; the recognition module will determine the identification data corresponding to the image data based on the vector weight and the image data and output it to the required terminal , The terminal can be, but not limited to, a smart phone, a tablet, a notebook computer, a desktop computer, or a smart wearable device.

Among them, in this creative content, the above-mentioned steps are included in the process of image pre-processing, cropping the image and text area that needs to be recognized.

According to an embodiment of the present creation, the method steps of the image recognition method include storing the recognition data as training data again.

According to an embodiment of the present creation, the method steps of the image recognition method during the training process further include comparing the recognition data with the training data, outputting a correlation coefficient α, and calculating the correlation Whether the number α is greater than a preset value K.

According to the content of this creation, the method steps of the image recognition method further include locating the graphic coordinates of the image data.

According to the content of this creation, the method steps of the image recognition method further include the graphic types of the classified image data.

According to the content of this creation, the method steps of the image recognition method further include correcting the angle of the image and text of the image data.

According to the content of this creation, the method steps of the image recognition method further include using the recognition module to convert the above-mentioned recognition data into corresponding characters or symbols.

The above descriptions are used to illustrate the purpose, technical means and achievable effects of this creation. Those who are familiar with this technology in the relevant field can use the following examples of demonstrations and accompanying schematic descriptions. The scope of the patent application makes this creation clearer.

100: Image recognition system

101: Control Module

103: Input Module

105: Image pre-processing module

105a: cutting unit

105c: correction unit

105e: Taxonomy unit

105g: positioning unit

107: Data Storage Module

109: Identification Module

109a: coding unit

109c: Decoding unit

201: image data

203: Decompose data

205: coded data

207: Decoding data

400: Image recognition method

S401-S407: Method steps

S501-S507: Method steps

The detailed description of this creation and the schematic diagram of the embodiments as described below should make this creation more fully understood; however, it should be understood that this is only used as a reference for understanding the application of this creation, and does not limit this creation to a specific implementation In the case.

Figure 1 illustrates the system architecture of the image recognition system.

Figure 2 shows how the encoding unit and the decoding unit process the input image data in the recognition module.

Figure 3A illustrates how the input image data is converted into encoded data by the image recognition system.

Figure 3B illustrates how the input image data is converted into encoded data by the image recognition system.

Figure 4 illustrates the method steps of the image recognition method.

Figure 5 further illustrates the steps of how to achieve image recognition through machine learning.

This creation will be described in detail with preferred embodiments and viewpoints. The following description provides specific implementation details of this creation, so that readers can thoroughly understand the implementation of these embodiments. However, those who are familiar with the art in this field must understand that this creation can also be implemented without these details. In addition, this creation can also be used and implemented by other specific embodiments. The details described in this specification can also be applied based on different needs, and various modifications or changes can be made without departing from the spirit of this creation. . This creation will be described with preferred embodiments and viewpoints. Such descriptions are used to explain the structure of this creation, and are only used to illustrate rather than limit the scope of patent application for this creation. In the method of this creation, the execution of each step The sequence may be executed before and after, or executed at the same time, which can be adjusted according to the needs of practical applications in the field. The terms used in the following description will be interpreted in the broadest reasonable way, even if they are used together with the detailed description of a specific embodiment of this creation.

The purpose of this creation is to try to improve the past graphic recognition technology. Although it can achieve a good recognition rate for images or texts with a fixed and neat format, it is scribbled or the recognized text has multiple expressions but represents the same When it is textual, the recognition rate will drop sharply. The technical means to improve the above-mentioned deficiency in this creation is to use machine learning to find out the optimized algorithm structure and the matching algorithm model, as well as the overall required parameter range, such as considering how to include in the limited memory In the case of body size, strike a balance between training and recognition efficiency, how to quickly converge to the best solution when the algorithm model extracts features of image data, how to adjust the weight of each feature extraction, and graphic data In order to achieve a better recognition rate, the graphic recognition technology can also be applied to more formal documents, such as the recognition of bank checks, contract documents, and legal documents. The further technical means of this creation will be detailed as follows Narrated.

In order to achieve the above objectives, in this creative strategy, in order for the image recognition system (100) to distinguish the input image data (201), the image recognition system (100) must first be trained by machine learning technology. A calculus model. The calculation model includes determining the coordinates, regions, angles, and types of images and texts that need to be interpreted in the image data (201) (for example, bank checks, contract documents, legal documents, etc.). If it is a bank check, it may be VIP Checks, non-VIP checks, or other types of checks), the type of graphic presentation (e.g., printing, handwriting), and the identification data of these graphic representations, (for example, "Five thousand and one round", " Five thousand and one yuan" represents the Arabic numeral "5001"). Since different coordinates, regions, angles, and types have been learned by the image recognition system (100) before the actual recognition of image data (201), the feature extraction and feature vector calculations will be more efficient, and the image data (201) can be accurately detected. ) To be identified in the data.

Therefore, based on the above strategy, please refer to Figure 1. This creation proposes an image recognition system (100). The system architecture includes: a control module (101) that provides system operation and management; and an input module (103) that inputs Image data (201); image pre-processing module (105) coupled to the input module (103) to adjust the data quality of the image data (201), wherein the image pre-processing module (105) includes a positioning unit (105g) , Locate the graphic coordinates of the image data (201); the recognition module (109), coupled to the image pre-processing module (105), decomposes the image data (201) into a plurality of decomposition data (203), by the positioning unit (105g) The location of the graphic coordinates, adjust the graphic area to be processed, and transmit the data coordinates back to the image pre-processing module (105), and output the identification data to the control module (101) according to the aforementioned graphic area ). Among them, in this creation, the control module (101) usually includes a processing chip, memory, display device, network communication module, operating system, application program, etc., which are connected to each other in a commonly known manner. Perform calculation, temporary storage, display and data transmission, and provide functions such as the operation and management coordination of the image recognition system (100). Based on the control module (101), it belongs to a generally known architecture, so it will not be repeated here.

In addition, in an embodiment of the present creation, the identification data can be manually compared or automatically fed back to the identification module (109) to further train the machine learning algorithm model and improve the accuracy of data identification . According to one of the viewpoints of this creation, in addition to the identification data, the graphic coordinates, graphic regions, graphic angles, graphic colors, and graphic types in this creation can also be fed back to the recognition module (109 ) To improve the cutting unit (105a), correction unit (105c), classification unit (105e), The accuracy of the positioning unit (105g).

Wherein, the aforementioned image recognition system (100) includes a data storage module (107) for storing training data input by the input module (103). In a preferred embodiment of the present creation, the training data source may include any type of identification data in graphics and text (for example, the aforementioned bank checks, contract documents, legal documents, etc., or in graphics and text The type of text, such as "Five thousand and one round", "Five thousand and one yuan", which represents the Arabic numeral "5001"), and has been marked with the graphic coordinates, graphic area, graphic angle, and graphic in advance. The image data (201) of text color and graphic type; or, the recognition module (109) will feed back each output recognition data to the recognition module (109) to further modify the aforementioned graphic coordinates and graphics. Text area, graphic angle, graphic color, and graphic type. That is, after each identification data is generated, the identification result is corrected (manually or automatically by the image recognition system 100), and it will be used to improve the accuracy of the next identification data identification.

Please refer to Figure 2, Figure 3A and Figure 3B. According to one viewpoint of this creation, in order to better correspond to "convert the image data (201) into the coded data (205) expressed in vector form, and then the coded data (205) through The feature extraction matrix calculates the decoded data (207), which is finally converted into the identification data by the recognition module (109). This problem in the form of "Sequence to Sequence", in the preferred embodiment of this creation, recognizes The module (109) adopts an Encoder-Decoder based training architecture. Therefore, the recognition module (109) includes system components of an encoding unit (109a) and a decoding unit (109c). Among them, the encoding unit (109a) encodes several decomposed data (203) into encoded data (205), and the encoding unit (109a) decodes the encoded data (205) into decoded data (207), the encoding unit (109a) The calculation models of the decoding unit (109c) and the decoding unit (109c) can be selected according to the needs of the application, such as Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and long- and short-term memory networks. (Long Short-Term Memory, LSTM), Bidirectional RNN (BiRNN), Gated Recurrent Unit (GRU), Attention, etc. For example, the recognition module can use a cyclic gated network as the encoding unit (109a), but the decoding unit (109c) uses a bidirectional recurrent neural network. Among them, the encoded data (205) and decoded data (207) described in this creation can be expressed as a high-dimensional vector. For the coded data (205), in an embodiment of the present creation, it can be expressed as a binary value matrix according to the grayscale color of each decomposition data (203). For example, the part with higher gray-scale color is represented as 1, and the gray-scale color The lower part is indicated as zero. Then in the graphic "5", the expression of the coded data (205) can be a matrix of u x v, where u and v are positive integers other than 0, and their size can be selected according to the needs of the application.

In view of the above, in the preferred embodiment of this case, the coding unit (109a) uses a convolutional neural network, and the decoding unit (109c) uses a long-term short-term memory network, and uses an attention model to enhance the features in the encoded data (205) The vector weight of the vector avoids the phenomenon that the feature vector of the recognition module may be weakened when the number of image data (201) and training data to be learned is large. In addition, according to one of the viewpoints of this creation, in order to better enable the encoding unit (109a) and decoding unit (109c) in this creation to be applied to the image recognition system (100), such as the creation of training data and image pre-image The judgment execution in the processing module (105) (such as pre-processing of graphic coordinates, graphic area, graphic angle, graphic color, graphic type, etc.), the recognition module (109) outputs the recognition based on the aforementioned data The process of data can achieve better training efficiency and recognition efficiency. The various parameters of the image recognition system (100) can be set as follows: Epoch (training times): 5-65; Batch Size (batch size): 10- 1024; Early Stop: 5-30; Learning Rate: 10 ^-2 -10 ^-6 . Among them, the setting of the above parameters has the advantage of achieving a balance between memory efficiency and memory capacity, and at the same time achieving recognition efficiency, feature extraction, and optimization of recognition rate, so that the image recognition system (100) can be applied Identification of official documents, such as bank checks, contract documents, or legal documents.

According to this creative content, the image pre-processing module (105) includes a cutting unit (105a), and the way of cutting the image data (201) is shown in FIG. 3A. In the embodiment of this creation, when there are multiple images in the image data (201) to be recognized, the cropping unit (105a) crops out according to the training data stored in the data storage module (107) To identify the appropriate size of the image and text area (such as field, range), to filter out the image area of each image and text in advance. In the preferred embodiment of this authoring, the whole picture and text is regarded as a complete input, and the cutting unit (105a) separates the picture and text area where the picture and text are located from the original picture according to the coordinates of the picture and text. Cut individual characters in different pictures and texts (for example, cut five, thousand, zero, one, circle, and whole respectively). In the process of establishing the calculation model, the whole picture and text are also used, such as "Five thousand and one round", so that the recognition module (109) can distinguish the overall text, rather than individual irrelevant text. . In addition, in another embodiment of this creation, the cutting unit (105a) can also be used according to the needs of the application. The characters should be cut separately to avoid the connection between the text and the text because the strokes are connected to each other, so that the image recognition system (100) will treat the two texts to be recognized as one image and text, and make subsequent recognition Module (109) has an error in recognition. For example, the two characters "五" and "千" are regarded as one character instead of two independent characters. After the cutting unit (105a) cuts the appropriate image and text area, the result of each time can be transmitted back to the recognition module (109) to further correct the image and text positioned by the positioning unit (105g) next time Area, to improve the positioning accuracy of the graphic area (manually or automatically corrected by the image recognition system 100).

According to an embodiment of the present creation, the cutting method of the above-mentioned cutting unit (105a) is called automatic positioning and identification. For a detailed implementation method, please refer to FIG. 3A. It can first cut out a part of the input image data (such as a check) according to the needs of the application, such as the upper half or the lower half. For example, in FIG. 3A, according to the normal position of a check, the cutting list unit (109a) cuts out a graphic area to be recognized, such as a denomination field. The specific method is to automatically locate the graphic coordinates ( x, y ) through the positioning unit (105g), and use the ( x, y ) as the reference, and cut the range from △x and △y to the range that needs to be identified. , For example, the horizontal axis is from x- △x to x + △x and the vertical axis is from y- △y to y + △y to cut out the graphic area of "Five thousand zero one rounded". Afterwards, the recognition module (109) performs the subsequent recognition process according to the position cut by the cutting unit (109a).

According to this creative content, the image pre-processing module (105) includes a correction unit (105c) for correcting the color of the image data (201), such as the intensity of the RGB three colors. In an embodiment of this creation, when the image and text in the image data (201) to be recognized has any signal strength in the RGB three colors is too strong, for example, when the image and text in "Five Thousand Zero One Round" is too strong, The red stamp in the background is too strong, which may affect the recognition of images and texts by the recognition module (109), the correction unit (105c) can convert the image data ( The red intensity of 201) is reduced to prevent the image data (201) from being identified incorrectly due to the strong background color. Preferably, after the correction unit (105c) completes the color correction of the image and text, the result of each time is transmitted back to the recognition module (109). According to the recognition result of the recognition module (109), it is combined with manual or automatic The correction is used to further correct the graphic color to be corrected next time by the correction unit (105c) to improve the accuracy of the correction.

According to another embodiment of the present invention, the correction unit (105c) can also be used to correct the angle of the image data (201). Please refer to FIG. 3A for the correction method. In the embodiment of this creation, when the need to identify When the image data (201) has an inclination angle of δ, the correction unit (105c) corrects the angle of the image and text in the image and text area according to the training data stored in the data storage module (107) to avoid the image The image and text in the data (201) have too large an inclination angle, which makes the stroke recognition error in the image and text. Preferably, after the correction unit (105c) finishes correcting the graphic area, the result of each time is transmitted back to the recognition module (109) to further correct the graphic and text corrected next time by the correction unit (105c) Area, improve the accuracy of the correction of the graphic area. Among them, in an embodiment of the present creation, the correction can be manual or automated.

According to this creative content, the image pre-processing module (105) includes a classification unit (105e). In the embodiment of this creation, before the image data (201) is recognized by the recognition module (109), the classification unit (105e) classifies the training data according to the training data stored in the data storage module (107) The document types in the image data (201) are bank checks, contract documents, or legal documents. For example, when the file of the image data (201) belongs to a bank check, the classification unit (105e) classifies the bank check as a VIP check, a non-VIP check, or other kinds of checks through the training data. Each classification result is transmitted back to the recognition module (109) and image pre-processing module (105), so that the cutting unit (105a), correction unit (105c), and positioning unit (105g) can be more accurate The cutting, correcting, and positioning of the graphic coordinates, graphic area, graphic angle, and graphic color can improve the accuracy of the image pre-processing module (105).

Therefore, based on the purpose and strategy of this creation, please refer to Figures 4 and 5. This creation proposes an image recognition method (400). The method steps include: in step (S405), the image data (201) is classified Text type; in step (S401), input training data through the input module (103), and the image data (201) to be recognized, and mark the training data corresponding to the image data (201); in step (S501) , The training data and image data (201) are input into the coding unit (109a) of the recognition module (109), and a feature extraction matrix is converted into vector form coding data (205); in step (S502) the coding data (205) Input the decoding unit (109c), and convert a feature extraction matrix into decoded data in vector form (207); and send the decoded data (207) to the recognition module (109) in step (S503); execute step (S406), using the identification module (109) to convert the decoded data into corresponding identification data.

Among them, in an embodiment of the creation, the image recognition method includes the step (S404) performed during the image preprocessing of the creation, that is, cropping out the image and text area that needs to be recognized in the image data (201); and In step (S407), the decoded data (207) and the identification data are stored as training data.

Among them, in a preferred embodiment of this creation, the encoding-decoding network uses a convolutional neural network for its encoding architecture, and a long- and short-term memory network for its decoding architecture, and uses an attention model to enhance the encoded data (205 The vector weight of the feature vector in) avoids the phenomenon that the feature vector of the recognition module may be weakened when the amount of training image data (201) is large.

In a preferred embodiment of the present creation, the method of comparing the image data (201) with the identification data is by comparing the identification data corresponding to the image data (201) with the training data in the step (S504), and Output a correlation coefficient α . Wherein, in step (S505), it is determined whether the correlation coefficient α is greater than a preset value K, if yes, step (S507) is executed to store the decoded data (207) as training data, and it is determined whether to accept, pass or It is to retain and store the vector weight. If not, in step (S506), perform step (S404) to readjust the feature vector weight and decide whether to stop or reject subsequent steps.

According to the content of this creation, the image recognition method (400) further includes performing step (S402), locating the graphic coordinates of the image data (201) by the positioning unit (105g); in step (S403), by correcting The unit (105c) corrects the angle and color of the image and text of the image data (201).

The above description is the preferred embodiment of this creation. Those skilled in this field should be able to understand that it is used to illustrate the creation and not to limit the scope of the patent rights claimed by this creation. The scope of its patent protection shall be determined by the scope of the attached patent application and its equivalent fields. Anyone familiar with the art in this field, without departing from the spirit or scope of this patent, makes changes or modifications that are equivalent changes or designs completed under the spirit of this creation, and should be included in the scope of the following patent application Inside.

100: Image recognition system

101: Control Module

103: Input Module

105: Image pre-processing module

105a: cutting unit

105c: correction unit

105e: Taxonomy unit

105g: positioning unit

107: Data Storage Module

109: Identification Module

109a: coding unit

109c: Decoding unit

Claims (10)

An image recognition system, including: A control module to provide system operation and management; An input module to input image data; An image pre-processing module coupled to the input module to adjust the data quality of the image data, wherein the image pre-processing module further includes a positioning unit for locating at least one graphic coordinate of the image data; and, An identification module, coupled to the image pre-processing module, decomposes the image data into a plurality of decomposition data, and recognizes and outputs at least one identification data according to the at least one graphic coordinate. The image recognition system according to claim 1, wherein the image pre-processing module further includes a cropping unit for cropping the image data into at least one graphic area according to at least one graphic coordinate positioned by the positioning unit , Thereby adjusting the range of processing required by the image recognition system. The image recognition system according to claim 1, wherein the image pre-processing module further includes a correction unit, and when the intensity of a color in the graphic data may affect the recognition of the recognition module, the correction unit adjusts the image data The color of a graphic in. The image recognition system according to claim 2, wherein the image pre-processing module further includes a classification unit for classifying the file types in the image data according to the at least one graphic area. The image recognition system according to claim 1, wherein the recognition module further includes an encoding unit for encoding the plurality of decomposition data into at least one encoded data in the form of a vector; and a decoding unit for encoding the at least one The encoded data is output as at least one decoded data. The image recognition system according to claim 5, wherein the calculation model of the encoding unit is a convolutional neural network; the calculation model of the decoding unit is a long and short-term memory network. The image recognition system according to claim 1, wherein the control module sets the number of times of training (Epoch) of the recognition module to 5-65 times. The image recognition system according to claim 1, wherein the control module sets the batch size of the recognition module to 10-1024. The image recognition system according to claim 1, wherein the control module sets the early stop value (Early Stop) of the recognition module to 5-30. The image recognition system according to claim 1, wherein the control module sets the learning rate of the recognition module to 10 ^-2 -10 ^-6 .

Images