Image category recognition method, readable storage medium and computer device

Abstract

The embodiment of the application provides a kind of image category identification method, readable storage medium and computer equipment, the present application still needs to extract the text image feature of seal local on the basis of existing identification algorithm, and the shape feature, color feature, spatial geometric feature extracted by existing algorithm are fused, such as the kind of court seal, the kind of invoice printing, the kind of passport entry and exit seal, including outbound or inbound, the kind of specific country, the true and false of identification seal, to facilitate the robustness and generalization performance of seal identification algorithm is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a kind of image kind identification method, readable storage medium and computer equipment. BACKGROUND

[0002] The shape feature, color feature, spatial geometric structure feature and local text image feature of seal need to be extracted.The existing deep learning identification algorithm can extract shape feature, color feature and spatial geometric structure feature, such as classification algorithm resnet, googlenet and SEnet on imagenet data set.

[0003] Most of the existing seal identification algorithms extract the text information on the seal through OCR algorithm.The existing OCR algorithm is divided into two kinds:one is two-stage:the first stage locates the text area through CNN text positioning algorithm, and the second stage locates the text area, and then recognizes the text in the located text area through RNN algorithm;The other kind of OCR algorithm is end-to-end, and simultaneously performs text positioning and text recognition.

[0004] The existing image recognition algorithms such as resnet, SEnet and googlenet can only extract shape feature, color feature and spatial geometric feature, which leads to the fact that the seal with only local text difference cannot be identified, and the seal with local text difference is common, so the existing algorithm has poor universality and generalization, and cannot extract local text image feature. SUMMARY

[0005] The purpose of the embodiment of the present application is to provide a kind of image kind identification method, readable storage medium and computer equipment.

[0006] In order to achieve the above-mentioned purpose, the first aspect of the present application provides a kind of image kind identification method, the method comprises:

[0007] input the image to be identified into image kind identification network, the image to be identified includes seal image;

[0008] obtain the first feature vector extracted by the first convolutional neural network of the image kind identification network on the image to be identified;

[0009] obtain the classification label image block of the image to be identified;

[0010] obtain the second feature vector extracted by the second convolutional neural network of the image kind identification network according to the classification label image block of the image to be identified on the image to be identified, and the second feature vector is the local feature of the character-containing seal image to be identified;

[0011] obtaining a third feature vector after fusing the first feature vector and the second feature vector;

[0012] performing recognition on the image to be recognized according to the third feature vector;

[0013] outputting the category of the image to be recognized.

[0014] In an embodiment of the present application, the step of obtaining the classification label image block of the image to be recognized comprises:

[0015] scanning the image to be recognized according to a preset window size to obtain a pixel block of the image to be recognized with the preset window size;

[0016] determining whether the pixel block contains characters;

[0017] if the pixel block contains characters, determining that the label of the pixel block is 1; if the pixel block does not contain characters, determining that the label of the pixel block is 0;

[0018] obtaining the classification label image block of the image to be recognized with the label value.

[0019] In an embodiment of the present application, the step of obtaining the second feature vector of the image to be recognized extracted by the second convolutional neural network of the image category recognition network according to the label of the image to be recognized comprises:

[0020] when the label of the pixel block is 0, the second convolutional neural network of the image category recognition network does not perform feature extraction on the image to be recognized;

[0021] when the label of the pixel block is 1, obtaining the local feature containing characters extracted by the second convolutional neural network of the image category recognition network from the image to be recognized.

[0022] In an embodiment of the present application, the step of determining whether the pixel block contains characters comprises:

[0023] obtaining positive example pixel blocks of images containing characters and negative example pixel blocks of non-character images;

[0024] inputting the positive example pixel blocks and the negative example pixel blocks into a classification convolutional neural network to output character image features through the classification convolutional neural network.

[0025] In an embodiment of the present application, the step of determining whether the pixel block contains characters further comprises:

[0026] inputting the classification label image block with the label value into a learning network;

[0027] Adaptively learning the classification label image block with the label value in the learning network through wx+b1 to determine the position and importance index of the character in the classification label image block;

[0028] wherein w is the label value of the classification label image block, x is the pixel of the classification image block, and b1 is an adaptive learning parameter.

[0029] In an embodiment of the present application, before the step of scanning the image to be recognized according to a preset window size to obtain a pixel block of the image to be recognized with the preset window size, the method further comprises:

[0030] Fixing the size of the image to be recognized as a preset size.

[0031] In an embodiment of the present application, the step of obtaining a third feature vector after the fusion processing of the first feature vector and the second feature vector comprises:

[0032] Splicing the first feature vector and the second feature vector;

[0033] Fusion processing the global feature of the spliced first feature vector and the local feature containing the character of the second feature vector;

[0034] Obtaining the third feature vector after the fusion processing.

[0035] In an embodiment of the present application, the step of fusion processing the global feature of the spliced first feature vector and the local feature containing the character of the second feature vector comprises:

[0036] Obtaining a feature vector after splicing the first feature vector and a plurality of second feature vectors;

[0037] Fusion processing the spliced feature vector y according to ay+b2;

[0038] Obtaining the third feature vector after the fusion processing.

[0039] In an embodiment of the present application, the preset window size is the size of one character in the image to be recognized.

[0040] The second aspect of the present application provides a processor configured to execute the above-mentioned image category identification method.

[0041] The third aspect of the present application provides a computer readable storage medium configured to execute the above-mentioned image category identification method.

[0042] The present application also needs to extract the local character image features of the seal on the basis of the existing recognition algorithm, and fuse the shape features, color features and spatial geometric features extracted by the existing algorithm, such as the type of court seal, the type of invoice printing, the type of passport entry and exit seal, including exit or entry, the type of specific country, the authenticity of the identification seal, etc., so as to improve the robustness and generalization performance of the seal recognition algorithm.

[0043] Other features and advantages of the embodiments of the present application will be described in detail in the following specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0044] The accompanying drawings are included to provide a further understanding of the embodiments of the present application, and constitute a part of the specification, and are used to explain the embodiments of the present application together with the following specific embodiments, but do not constitute a limitation on the embodiments of the present application. In the drawings:

[0045] Figure 1 The flowchart of the method for image type recognition according to the embodiments of the present application is schematically shown;

[0046] Figure 2 The pixel block example diagram of the image to be recognized according to the embodiments of the present application is schematically shown;

[0047] Figure 3 The internal structure diagram of the computer device according to the embodiments of the present application is schematically shown. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. It should be understood that the specific embodiments described here are only used to illustrate and explain the embodiments of the present application, and are not used to limit the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0049] Specifically, the existing convolutional network architecture, from the characteristics of convolution, the sampling grid of convolution is fixed, such as 3x3, and the parameters of convolution are shared, that is, the parameters of the filter are the same for local text images and non-local text images, so the extracted features are the same, and therefore local text image features cannot be extracted. Moreover, the spatial sampling point position of convolution is fixed, such as a grid, and the text on the seal is mostly irregular, such as on an arc, and curved, so it is more difficult to extract such curved local text features. In summary, the characteristics of convolution can be used to identify seals by shape, color, and spatial geometric features, which makes it difficult to identify seals that only differ in local text, and local text different seals are common, so the existing algorithm has poor generality and generalization. Secondly, representing the differences between seal categories and mining differentiated image features, the extraction of local text image features is crucial. Therefore, the present embodiment proposes the following method to solve the above problems.

[0050] Figure 1 The flowchart of the method for image category recognition according to the embodiments of the present application is schematically shown. As shown in Figure 1 In an embodiment of the present application, a method for image category recognition is provided, comprising the following steps:

[0051] Step S10, inputting a to-be-identified image into an image category recognition network, wherein the to-be-identified image comprises a seal image;

[0052] Step S20, obtaining a first feature vector extracted by a first convolutional neural network of the image category recognition network from the to-be-identified image;

[0053] Step S30, obtaining a classification label image block of the to-be-identified image;

[0054] Step S40, obtaining a second feature vector extracted by a second convolutional neural network of the image category recognition network from the to-be-identified image according to the label of the to-be-identified image;

[0055] Step S50, obtaining a third feature vector obtained by performing fusion processing on the first feature vector and the second feature vector;

[0056] Step S60, identifying the to-be-identified image according to the third feature vector;

[0057] Step S70, outputting the category of the to-be-identified image.

[0058] The processing method is proposed for the scenario of identifying images, specifically, identifying text information in images to distinguish the categories of the real objects corresponding to the images. In order to facilitate the description, the method of the present embodiment is described by taking a seal as an example:

[0059] The image to be identified is input into the image category identification model, and the first convolutional neural network of the image category identification model is used to extract features of the tourism seal image to be identified. Specifically, the shape feature, color feature, spatial geometric structure feature, and global feature of the seal are extracted through the first convolutional neural network to obtain a first feature vector. The spatial geometric structure feature of the seal is the context position relationship, pixel relationship, texture relationship, and geometric relationship on the seal. The global feature of the seal is an abstract feature that can identify the identity ID of the seal. Specifically, for entry and exit seals, the spatial geometric structure feature is the position and relationship of the text area and non-text area on the entry and exit seal, the combination relationship of text pixels and non-text pixels, the combination relationship of text texture and non-text texture, and the combination relationship of the outline shape surrounded by the text and the peripheral outline shape of the seal. The global feature is the identity of the object, such as the object identity of this type of seal is the entry seal of China, such as the face picture of a person in face recognition identifies the identity ID of the person or the category information of the person.

[0060] The first convolutional neural network structure can be resnet series, ResNeXt, DenseNet, mobilenet series, shufflenet, googlenet, SENet, mobilefacenet, or efficientnet. Through the first convolutional neural network structure, a feature vector with a predetermined dimension is obtained, which can be 64-dimensional, 128-dimensional, or 256-dimensional. The dimension size can be determined according to the number of seal categories and the inter-class difference. If different categories only have slight differences, a larger dimension can be used. If the inter-class difference is small, a smaller dimension can be used. When the number of seal categories increases, the difficulty of the classification task increases, and the dimension of the feature vector needs to be increased. When the differences between different categories of seals are small, it is easy to misidentify and recognize different categories of seals as the same category, increasing the difficulty of recognition and also requiring an increase in the dimension of the feature vector. Here, the greater the difficulty of recognition, the greater the dimension of the feature, so that the feature vector can express enough information, because higher discrimination ability is needed to distinguish between difficult-to-distinguish sample categories. Samples that cannot be distinguished in a low-dimensional feature space can be distinguished in a high-dimensional feature space.

[0061] The number of seal categories and the inter-class difference are obtained through business scenario data, such as the number of passport entry and exit seals and the difference between different national seals, which is obtained from data pixel distribution. A larger dimension is used for a smaller inter-class difference, and a smaller dimension is used for a larger inter-class difference. The category difference and the inter-class difference are the same concept. The difference size is measured according to the similarity between samples. For example, if the RGB pixel difference between two categories of images is small, the similarity is large, and the inter-class difference is small.

[0062] After the processor acquires the seal image to be recognized, the seal image is subjected to label adding and classification processing. The label adding and classification referred to herein means that the seal image is divided into blocks by a specific means, some of the blocks have characters and some do not have characters, and the seal image is divided into blocks, and local feature extraction is performed on the blocks with characters, and no feature extraction is performed on the blocks without characters. The specific method includes: scanning the image to be recognized according to a preset window size to obtain a pixel block of the image to be recognized with the preset window size; determining whether the pixel block contains characters; if the pixel block contains characters, determining that the label of the pixel block is 1; if the pixel block does not contain characters, determining that the label of the pixel block is 0; and obtaining a classification label image block of the image to be recognized with the label value.

[0063] The adjusted seal image is divided into a fixed size, and the adjusted seal image is scanned from top to bottom and from left to right through a window with a predetermined size to obtain a pixel block of the window with the predetermined size, as shown in FIG. 3. Figure 2 The standard of the pixel block is that the size of the pixel block is the average size of the single character pixel block, and the adjacent pixel blocks can also overlap by several pixels to improve the processing capability of the edge information of the adjacent pixel blocks and improve the expression capability of the features of the character pixel region.

[0064] The reason why the seal image to be recognized is adjusted to a fixed size is that a full connection layer will be used in the subsequent steps, and the full connection layer requires that the input is of a fixed size. In addition, the uniform image size is also to accelerate the calculation and improve the speed. If the image sizes are different, the corresponding convolution calculation amount is different, and the GEMM principle of convolution calculation cannot be applied to accelerate the calculation. The fixed size of the seal image in the method of the embodiment can be set according to the needs.

[0065] In the embodiment, the size range of the preset window size is the pixel size range of a single character. The pixel block is subjected to classification recognition to determine whether there are characters in the pixel block, and if there are, the pixel block is marked as 1, and if there are not, the pixel block is marked as 0. Finally, a classification label image block of the seal image is obtained.

[0066] In the embodiment, the method for determining whether there is a character in the pixel block comprises: obtaining positive example pixel blocks of a character-containing image and negative example pixel blocks of a non-character image; inputting the positive example pixel blocks and the negative example pixel blocks into a classification convolutional neural network, and outputting character image features through the classification convolutional neural network. Specifically, from public handwritten character images and arbitrary scene printed character images, pixel blocks of a preset window size are cut as positive examples, and from non-character images, pixel blocks of a preset window size are cut as negative examples. The difference features of character pixels and non-character pixels are extracted through a binary classification convolutional neural network structure, so as to realize the extraction of character image features and the distinction of background images. The training method of the binary classification convolutional neural network comprises: from public handwritten character images and arbitrary scene printed character images, pixel blocks of a preset window size are cut as positive examples, and from non-character images, pixel blocks of a preset window size are cut as negative examples. The positive example pixel blocks and the negative example pixel blocks are input into the binary classification convolutional neural network for training, and the difference features of the positive example pixels and the negative example pixels are extracted; and character image features are extracted according to the difference features of the positive example pixels and the negative example pixels. In this way, it is determined whether the pixel block of the image to be recognized contains a character.

[0067] Further, the local character image features of the classification label image block of the seal image are extracted through a second convolutional neural network structure, and a second feature vector is obtained.

[0068] The second convolutional neural network structure can be the same as the first convolutional neural network structure but with different parameters, or can be different from the first convolutional neural network structure and can be an efficient net or a reset series. Through the second convolutional neural network structure, a feature vector of a predetermined size is obtained, and the dimension of the feature vector can be 64, 128 or 256.

[0069] The classification label image block of the seal image obtained above is annotated with character pixel positions, with 1 for a character and 0 for no character, and the pixel positions of the classification label image block correspond one by one to the original seal pixel positions. Therefore, the classification label image block plays a switching role, and determines which pixels in the original seal pixels perform dot multiplication operation with the filters in the second convolutional neural network structure. The pixels at the corresponding positions of the seal are set to 0 in the classification label image block without a character, and after the dot multiplication operation with the filters, the pixels are still 0, that is, they do not participate in feature extraction. The pixels at the corresponding positions of the seal are unchanged in the classification label image block with a character, and the normal dot multiplication operation with the filters is performed to extract features. In this way, the network focusing ability is achieved at the positions with characters, and the local character image features are extracted.

[0070] In addition to the above method, the classification label image block of the seal image can also be obtained by the following method: firstly, coarse classification is performed by a weak supervision method, and then adaptive learning is performed on the coarsely classified image. The weak supervision method is as follows:

[0071] Method for determining whether there is a character in a pixel block:

[0072] From the public handwritten character image and the arbitrary scene printed character image, a pixel block of a preset window size is cut as a positive example, and a pixel block of a preset window size is cut from a non-character image as a negative example. The difference features of the character pixels and the non-character pixels are extracted by a binary classification convolutional neural network structure to realize the extraction of the character image features and the distinction of the background image.

[0073] Further, the classification label image block (containing values 0 and 1) obtained by coarse classification is used as an initial value, and the role of a similar FC layer is realized, that is, the input x, that is, the original seal input pixel, is affected by w in wx+b1. Here, w is the value of the classification label image block. Through network back propagation, the value in the classification label image block is learned to adaptively learn the position and importance of the character in the seal (the value of the classification label image block is less than 0, and the value of the classification label image block is greater than 0, which is represented as a character pixel. The greater the value, the greater the importance). Through the adaptive learning method of the classification label image block, the position of the seal character pixel can be more fine, and the background noise is more robust.

[0074] The original seal input pixel here is the original seal image corresponding to the classification label image block, that is, the seal image of a fixed size resized in step 2, that is, the input image used to obtain the classification label image block,

[0075] Here, b1 is a learnable parameter, which is a bias term, used to learn the deviation of the classification label image block from the real text region. Because the pixel block may not completely match the text region, it may predict the background noise as the text region. Here, a parameter b is added to alleviate this problem

[0076] The step of obtaining the third feature vector after fusing the first feature vector and the second feature vector in the above step includes: splicing the first feature vector and the second feature vector; fusing the global features and the local text image features of the spliced first feature vector and second feature vector; and obtaining the third feature vector after fusion processing.

[0077] The processor splices the first feature vector with the second feature vector, fuses global features such as seal shape features, color features, spatial geometric features and local text image features through a full connection layer, and obtains a final feature vector. In the feature splicing, the first feature vector is directly connected to the second feature vector, for example, the first feature vector is 64-dimensional, and the second feature vector is 64-dimensional, so that the spliced feature vector is 128-dimensional. The fusion operation is a w-point multiplication operation on the spliced feature vector y plus b2, so that the values in the feature vector y are correlated and fused.

[0078] It is to be noted that y in the above is the feature vector after splicing the first feature vector and the second feature vector, and the correlation between the first feature vector and the second feature vector is learned by multiplying w and adding b2. The dimensions of the original y feature vector after splicing are independent, that is, the first feature vector and the second feature vector are unrelated. The weight value w can be large or small, so that the importance of the values in each dimension of y is different, that is, there is an attention mechanism. The importance of each component of the text feature is not the same, the importance of each component of the global feature and the spatial geometric feature is also different, and the importance of each component of the feature after combining the local text feature and the global feature is also different. For example, the surrounding text is used to infer the local text, and the role or importance of the surrounding text is not the same.

[0079] The final feature vector fuses the local text image features in the second feature vector and the seal shape features, color features and spatial geometric features in the first feature vector, represents the differences between seal categories, and excavates differentiated image features. The final feature vector has stronger discriminative ability and generalization performance. In addition, the final feature vector is learned according to the task. For example, the features required for identifying the entry and exit seals of a certain country include shape, color, text content, spatial geometric structure and global features. The weight proportion of various features, that is, which features are combined to play a role in classification, is a parameter to be learned by the model.

[0080] The present application further extracts local text image features of the seal on the basis of the existing recognition algorithm, and fuses the shape features, color features and spatial geometric features extracted by the existing algorithm. For example, the present application can identify the categories of court seals, invoice seals, entry and exit seals of passports, categories of specific countries, and authenticity of seals, so as to improve the robustness and generalization performance of the seal recognition algorithm.

[0081] The embodiment of the present application provides a storage medium having a program stored thereon, and the program is executed by a processor to implement the image category recognition method.

[0082] The embodiment of the present application provides a processor, which is used for running a program, wherein the program performs the image category identification method.

[0083] In one embodiment, a computer device, which can be a server, has an internal structure diagram as shown in the figure. Figure 3 The computer device includes a processor A01, a network interface A02, a memory (not shown in the figure) and a database (not shown in the figure) connected through a system bus. The processor A01 of the computer device is used for providing computing and control capabilities. The memory of the computer device includes an internal memory A03 and a non-volatile storage medium A04. The non-volatile storage medium A04 stores an operating system B01, a computer program B02 and a database (not shown in the figure). The internal memory A03 provides an environment for the operating system B01 and the computer program B02 in the non-volatile storage medium A04. The database of the computer device is used for storing working condition data, working condition parameters and the like. The network interface A02 of the computer device is used for communicating with an external terminal through a network connection. The computer program B02 is executed by the processor A01 to implement the image category identification method.

[0084] Those skilled in the art can understand that, Figure 3 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0085] The embodiment of the present application provides a device, which includes a processor, a memory and a program stored in the memory and executable on the processor. When the processor executes the program, the following steps are implemented: inputting an image to be identified into an image category identification network; obtaining a first feature vector extracted by a first convolutional neural network of the image category identification network from the image to be identified; obtaining a classification label image block of the image to be identified; obtaining a second feature vector extracted by a second convolutional neural network of the image category identification network from the image to be identified according to the classification label image block of the image to be identified; obtaining a third feature vector obtained by performing fusion processing on the first feature vector and the second feature vector; identifying the image to be identified according to the third feature vector; and outputting a category of the image to be identified.

[0086] In an embodiment, the step of obtaining the classification label image block of the image to be recognized includes: scanning the image to be recognized according to a preset window size to obtain a pixel block of the image to be recognized with the preset window size; determining whether the pixel block contains characters; if the pixel block contains characters, determining that the label of the pixel block is 1; if the pixel block does not contain characters, determining that the label of the pixel block is 0; and obtaining the classification label image block of the image to be recognized with the label value.

[0087] In an embodiment, the step of extracting, by the second convolutional neural network of the image category recognition network, the second feature vector of the image to be recognized according to the label of the image to be recognized includes: when the label of the pixel block is 0, the second convolutional neural network of the image category recognition network does not extract features of the image to be recognized; and when the label of the pixel block is 1, obtaining the local character image features extracted by the second convolutional neural network of the image category recognition network from the image to be recognized.

[0088] In an embodiment, the step of determining whether the pixel block contains characters includes: obtaining positive example pixel blocks of images containing characters and negative example pixel blocks of non-character images; extracting difference features of the positive example pixels and the negative example pixels; and determining whether the pixel block of the image to be recognized contains characters according to the difference features of the positive example pixels and the negative example pixels.

[0089] In an embodiment, the step of determining whether the pixel block contains characters further includes: inputting the classification label image block with the label value into a learning network; and performing adaptive learning on the classification label image block with the label value in the learning network by wx+b1 to determine the position and importance index of characters in the classification label image block; wherein w is the label value of the classification label image block, x is the pixel of the classification image block, and b1 is an adaptive learning parameter.

[0090] In an embodiment, before the step of scanning the image to be recognized according to a preset window size to obtain a pixel block of the image to be recognized with the preset window size, the method further includes: fixing the size of the image to be recognized to a preset size.

[0091] In an embodiment, the step of obtaining the third feature vector obtained by fusing the first feature vector and the second feature vector includes: splicing the first feature vector and the second feature vector; fusing the global features and the local character image features of the spliced first feature vector and second feature vector; and obtaining the third feature vector after the fusion processing.

[0092] In one embodiment, the step of fusing the global feature and the local character image feature of the spliced first feature vector and the second feature vector comprises: processing the spliced feature vector y according to ay+b2; and obtaining a fused third feature vector.

[0093] In one embodiment, the preset window size is the size of a character in the image to be recognized.

[0094] The present application also provides a computer program product adapted to execute the program steps of the above method when executed on a data processing device.

[0095] Those skilled in the art will understand that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) having computer-usable program code embodied therein.

[0096] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions, which are executed via the processor of the computer or other programmable data processing device, generate a device implemented in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 means for performing the function specified by the flow or flows and / or block or blocks.

[0097] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction means, which implement the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 means for performing the function specified by the flow or flows and / or block or blocks.

[0098] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are performed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide a process for implementing the flowcharts and / or block diagrams. Figure 1one or more processes and / or blocks Figure 1 steps of a function specified in one or more blocks.

[0099] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0100] The memory can include non-persistent memory and / or persistent memory, both of which can be volatile and / or non-volatile. Non-volatile storage is typically persistent unless otherwise indicated. The memory is an example of computer readable media.

[0101] Computer readable media includes permanent and non-permanent, moveable and non-moveable media that can be implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile discs (DVDs) or other optical storage, magnetic cassettes, magnetic tapes, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible to a computing device. According to the definition herein, computer readable media does not include transitory media, such as modulated data signals and carrier waves.

[0102] It should also be noted that the terms "comprising," "including," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without limitation, an element preceded by "comprises a" does not, without more constraints, foreclose the existence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0103] The above merely provides an example of the present application and is not intended to limit the present application. The present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application should be included in the scope of claims of the present application.

Claims

1. A method of image category recognition, characterized by, The method comprises: inputting an image to be identified into an image category identification network, the image to be identified comprising a seal image; obtaining a first feature vector extracted by a first convolutional neural network of the image category identification network from the image to be identified; obtaining a classification label image block of the image to be identified with a label value, the label value indicating whether a pixel block contains a character; obtaining a second feature vector extracted by a second convolutional neural network of the image category identification network from the image to be identified according to the classification label image block of the image to be identified, the second feature vector being a local feature of the seal image to be identified containing a character; obtaining a third feature vector obtained by fusing the first feature vector and the second feature vector; identifying the image to be identified according to the third feature vector; outputting a category of the image to be identified.

2. The method of claim 1, wherein, The step of obtaining the classification label image block of the image to be identified comprises: scanning the image to be identified according to a preset window size to obtain a pixel block of the image to be identified with the preset window size; determining whether the pixel block contains a character; if the pixel block contains a character, determining that a label of the pixel block is 1; if the pixel block does not contain a character, determining that the label of the pixel block is 0; obtaining the classification label image block of the image to be identified with the label value.

3. The method of claim 2, wherein, The step of obtaining the second feature vector extracted by the second convolutional neural network of the image category identification network from the image to be identified according to the classification label image block of the image to be identified comprises: when the label of the pixel block is 0, the second convolutional neural network of the image category identification network does not extract a feature from the image to be identified; when the label of the pixel block is 1, obtaining a local feature containing a character extracted by the second convolutional neural network of the image category identification network from the image to be identified.

4. The method of claim 2, wherein, The step of determining whether the pixel block contains a character comprises: obtaining a positive example pixel block of a character-containing image and a negative example pixel block of a non-character image; inputting the positive example pixel block and the negative example pixel block into a classification convolutional neural network to output a character image feature through the classification convolutional neural network.

5. The method of claim 2, wherein, The step of determining whether the pixel block contains a character further comprises: inputting the classification label image block with the label value into a learning network; performing adaptive learning on the classification label image block with the label value in the learning network through wx+b1 to determine a position and an importance index of a character in the classification label image block; wherein w is a label value of the classification label image block, x is a pixel of the classification image block, and b1 is an adaptive learning parameter.

6. The method of claim 2, wherein, Before the step of scanning the image to be identified according to a preset window size to obtain a pixel block of the image to be identified with the preset window size, the method further comprises: fixing a size of the image to be identified as a preset size.

7. The method of claim 1, wherein, The step of obtaining the third feature vector obtained by fusing the first feature vector and the second feature vector comprises: splicing the first feature vector and the second feature vector; fuse the global feature of the first feature vector after splicing and the local feature containing the character of the second feature vector; obtain the third feature vector after fusion processing.

8. The method of claim 7, wherein, The step of fusing the global feature of the first feature vector after splicing and the local feature containing the character of the second feature vector comprises: obtain the feature vector after splicing the first feature vector and the majority of the second feature vectors; fuse the feature vector y after splicing according to ay+b2; obtain the third feature vector after fusion.

9. A computer-readable storage medium, characterized in that, An apparatus configured to perform the method of image category recognition according to any one of claims 1 to 8.

10. A computer device, comprising: An apparatus configured to perform the method of image category recognition according to any one of claims 1 to 8.

Images