A text-independent off-line writer retrieval method, device and storage medium

Abstract

The application discloses a text-independent handwriting retriever, a device and a storage medium, and belongs to the field of deep learning and pattern recognition. The method comprises the following steps: obtaining text-independent handwriting time sequence data; performing normalization processing on the obtained handwriting time sequence data, extracting a plurality of time functions, and obtaining label data; constructing a retrieval model, training the retrieval model according to the normalized time sequence data and the label data; inputting test handwriting and all test handwriting in a database into the trained retrieval model, and extracting a feature vector; calculating the cosine similarity between the feature vector of the test handwriting and the feature vector of the test handwriting in the database, and obtaining the final retrieval result according to the cosine similarity. The application aims at online writer retrieval, fills the long-term research gap in the field, extracts the writing style of the same writer in different texts, and achieves retrieval performance far exceeding other methods.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of deep learning and pattern recognition, and particularly relates to a text-independent online writer retrieval method and device and a storage medium. BACKGROUND

[0002] Handwriting is very common in daily life, such as handwritten notes, handwritten signatures, etc. Handwriting contains rich personal writing features and can be used for identity recognition, and is widely used in criminal investigation, finance and other fields. Since entering the digital era, due to the maturity of touch screen devices and electronic stylus, online (online) handwriting applications have been popular in various fields, such as conference demonstrations, electronic reading notes, etc., generating a wealth of online handwriting data. Therefore, retrieving the corresponding writer from online handwriting, i.e. online writer retrieval, shows important application value, such as finding suspects in criminal investigation, identity authentication in financial services, etc.

[0003] In recent years, although offline writer retrieval is developing rapidly, online writer retrieval has been neglected, and there is currently a lack of corresponding solutions. SUMMARY

[0004] To at least partially solve one of the technical problems existing in the prior art, the purpose of the present application is to provide a text-independent online writer retrieval method, device and storage medium based on time-frequency collaborative representation learning.

[0005] The first technical solution adopted by the present application is:

[0006] A text-independent online writer retrieval method, comprising the following steps:

[0007] Obtaining text-independent handwriting time series data;

[0008] Normalizing the obtained handwriting time series data and extracting a plurality of time functions to obtain label data;

[0009] Constructing a retrieval model and training the retrieval model according to the normalized time series data and the label data;

[0010] Inputting the test handwriting and all test handwriting in the database into the trained retrieval model to extract feature vectors;

[0011] Calculating the cosine similarity between the feature vectors of the test handwriting and the feature vectors of the test handwriting in the database, and obtaining the final retrieval result according to the cosine similarity.

[0012] Further, the normalization processing includes pressure normalization processing and coordinate center normalization processing.

[0013] Further, the time function comprises at least one of a first derivative of the x coordinate, a first derivative of the y coordinate, a second derivative of the x coordinate, a second derivative of the y coordinate, a velocity, a tangent angle between two points, a cos value of the tangent angle, a sin value of the tangent angle, a first derivative of the velocity, a first derivative of the tangent angle, a logarithmic curvature radius, a velocity variation amplitude, an acceleration, or a pressure.

[0014] Further, the retrieval model is a deep neural network model, and the retrieval model comprises a time domain redundancy reduction subnetwork, a frequency domain cross-attention subnetwork, and a context-aware feature pyramid network.

[0015] The time domain redundancy reduction subnetwork is composed of a plurality of channel activation reverse residual modules, and is configured to reduce redundancy in a time domain feature channel dimension through multi-level channel activation.

[0016] The frequency domain cross-attention subnetwork is configured to extract a high-frequency component of a time sequence through discrete wavelet transform, and perform gated cross-attention calculation on the time sequence and the high-frequency component to highlight writing style features of the high-frequency component.

[0017] The context-aware feature pyramid network is configured to obtain multi-scale outputs of the time domain redundancy reduction subnetwork and perform context enhancement, and output a time domain feature vector in a pyramid structure.

[0018] Further, the output of the frequency domain cross-attention subnetwork is subjected to pooling to obtain a frequency domain feature vector.

[0019] In training the model, both the time domain feature vector and the frequency domain feature vector are used in training; in testing the model, only the time domain feature vector is used for inter-sample matching.

[0020] Further, the channel activation reverse residual module works as follows:

[0021] The input is evenly divided into two tensors according to the channel dimension, one of which is subjected to a reverse residual module, mapped from d dimensions to d dimensions, then activated by a squeeze-and-excitation layer, and finally mapped to d dimensions through a convolution-nonlinear layer. The other tensor is directly mapped from d dimensions to d dimensions through a convolution layer for mapping; the two tensors are spliced according to the channel dimension to obtain a tensor with a shape of l x o; wherein c is the input channel dimension, l is the sequence length, d is the intermediate layer channel dimension, and o is the output channel dimension.

[0022] The channel is activated again, and finally output through a post-processing linear layer and a dropout layer. ​​​

[0023] Furthermore, the frequency domain cross-attention sub-network operates as follows:

[0024] The frequency-domain cross-attention subnetwork receives the output of the first layer of the time-domain redundancy removal subnetwork as input. It extracts high-frequency components using discrete wavelet transform, then performs cross-attention on the high-frequency components and the original time features. A tanh(a) value is then used as a gate to multiply the cross-attention result. a is initialized to 0 and will be optimized with gradient updates. The output of this frequency-domain cross-attention subnetwork is fed into a selective pooling layer, outputting a frequency-domain feature vector f. F ;

[0025] In addition, the frequency domain cross-attention subnetwork and the time domain redundancy removal subnetwork perform feature interaction between layers. That is, the input of the next time domain layer is added to the output of the previous frequency domain layer, and the input of the next frequency domain layer is added to the output of the current time domain layer, forming a collaborative feature learning between the time domain and the frequency domain.

[0026] Furthermore, the context-focused feature pyramid network operates as follows:

[0027] The context-focused feature pyramid network consists of two pyramid layers. Each pyramid layer receives the outputs of the two-scale temporal redundancy-removal subnetwork layers as inputs. The two inputs are fed into the context building module for context modeling. Then, one feature is upsampled and added to the feature at the other scale. After addition, it is compressed into a feature vector by a selective pooling module to complete the calculation of one pyramid layer.

[0028] The temporal features of the two-layer pyramid outputs are f T1 and f T2 These two time-domain features are concatenated along the channel dimension to obtain the time-domain feature f. T .

[0029] The second technical solution adopted in this invention is:

[0030] A text-based autologous writer retrieval device, comprising:

[0031] At least one processor;

[0032] At least one memory for storing at least one program;

[0033] When the at least one program is executed by the at least one processor, the at least one processor implements the method described above.

[0034] The third technical solution adopted in this invention is:

[0035] A computer-readable storage medium having stored therein a program executable by a processor for performing the method as described above when executed by the processor.

[0036] The present application has the beneficial effect that the present application fills the long research gap in online writer retrieval, and can enhance the mining of writing style in online handwriting through time-frequency feature collaborative learning, extract the writing style of the same writer in different texts, and achieve retrieval performance far superior to other methods. BRIEF DESCRIPTION OF DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following introduces the drawings of the related technical solutions in the embodiments of the present application or the prior art. It should be understood that the drawings in the following introduction are only for the convenience of clearly describing part of the embodiments in the technical solutions of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the premise that there is no creative labor.

[0038] Figure 1 is a flowchart of a text-independent online writer retrieval method in the embodiments of the present application;

[0039] Figure 2 is a structure diagram of a retrieval model in the embodiments of the present application. DETAILED DESCRIPTION

[0040] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the drawings, wherein the same or similar reference signs represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the drawings are exemplary and are only used to explain the present application, and cannot be understood as a limitation of the present application. For the step numbers in the following embodiments, they are only set for the convenience of description and explanation, and the order between the steps is not limited in any way, and the execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0041] In the description of the present application, it should be understood that the orientation description, such as the orientation or position relationship indicated by up, down, front, back, left, right, etc. is based on the orientation or position relationship shown in the drawings, and is only for the convenience of describing the present application and simplifying the description, and does not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation of the present application.

[0042] In the description of the present application, the meaning of one or more is one or more, the meaning of multiple is two or more, greater than, less than, more than, etc. are understood as not including the number, above, below, within, etc. are understood as including the number. If it is described that the first, second is only used for the purpose of distinguishing technical features, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features or implicitly indicating the order of the indicated technical features. In addition, "and / or" describes the association relationship of the associated objects, which means that there can be three kinds of relationships, for example, A and / or B can represent: A exists alone, A and B exist together, and B exists alone. The character " / " generally represents an "or" relationship between the front and rear associated objects.

[0043] In the description of the present application, unless otherwise explicitly limited, the words such as setting, installing, connecting, etc. should be broadly understood, and those skilled in the art can reasonably determine the specific meaning of the above words in the present application in combination with the specific content of the technical solution.

[0044] In view of the existing technical problems and from the practical point of view, the current writer retrieval involves a text-independent scene, that is, any content written by the writer should be retrieved. Therefore, the present application proposes a representation learning method based on deep learning and time-frequency cooperation for text-independent online writer retrieval. The method can remove redundant features in the time domain through multi-level channel activation, extract refined time features, then enhance and extract high-frequency components in the frequency domain through the gating cross-attention mechanism and discrete wavelet transform, and finally realize time-frequency cooperation by interacting time and frequency features. This method can extract robust handwriting features in a text-independent scene for accurate retrieval matching.

[0045] As Figure 1 described, the embodiment provides a text-independent online writer retrieval method, comprising the following steps:

[0046] S1, obtaining text-independent handwriting time sequence data.

[0047] Data preparation: obtaining text-independent handwriting time sequence data. This embodiment uses random number strings with different contents as data examples, and other text-independent handwriting data can be used. The storage format of the random number string is time sequence, which contains the x, y coordinates and pressure information of each handwriting point.

[0048] S2, normalizing the obtained handwriting time sequence data, and extracting a plurality of time functions to obtain label data.

[0049] Data preprocessing: all data is pressure normalized and coordinate center normalized, and 14 time functions are extracted. As an optional implementation, step S2 specifically includes the following steps:

[0050] S21, normalizing the pressure information of the data.

[0051] S22, center normalizing the original x, y coordinate information of the data, that is, taking the midpoint of the sequence coordinates as the origin to establish a new coordinate system, and compressing other coordinates to the interval (-1, 1) in proportion, and mapping to the new coordinate system.

[0052] S23, using x, y coordinates and pressure to extract 14 time functions, as shown in Table 1 below:

[0053] Table 1 Time function list

[0054]

[0055] S3, constructing a retrieval model, training the retrieval model according to the normalized time series data and label data.

[0056] In some optional embodiments, a deep neural network model is constructed as the retrieval model, which includes a time domain redundancy reduction subnetwork, a frequency domain cross attention subnetwork, and a context attention feature pyramid network. The time domain redundancy reduction subnetwork is composed of a channel activation anti-residual module, which reduces the redundancy in the channel dimension of the time domain features through multi-level channel activation. The frequency domain cross attention subnetwork extracts the high frequency components of the time series through discrete wavelet transform, and performs gated cross attention calculation on the time series and its high frequency part to highlight the writing style features of the high frequency part. The context attention feature pyramid network obtains the multi-scale output of the time domain redundancy reduction subnetwork and performs context enhancement, and outputs the time domain feature vector according to the structure of the pyramid. The output of the frequency domain subnetwork is directly obtained after pooling to obtain the frequency domain feature vector. Both the time domain feature vector and the frequency domain feature vector are used in training. When testing, the model only uses the time domain feature vector for inter-sample matching. The network model structure diagram is as shown in Figure 2 .

[0057] In some optional embodiments, the structure of the time domain redundancy reduction subnetwork is as shown in Table 2, and Table 3 lists the structure of some sublayers thereof respectively, and l is the length of the input handwriting sequence. In order to facilitate representation, the input length of each table below is represented by l.

[0058] Table 2 Time domain redundancy reduction subnetwork structure

[0059]

[0060]

[0061] Table 3 Channel activation inverse residual module structure

[0062]

[0063] Specifically, for each channel activation inverse residual module, the input is first split into two halves along the channel dimension. One half passes through the inverse residual module, which is composed of a normal convolution and a depthwise separable convolution, from dimensional mapping to d-dimensional, and then passes through a squeeze-and-excitation layer to activate the channels, and finally maps to dimensional through a convolution-nonlinear layer. The other half directly passes through a convolution layer for mapping from to dimensional. Then the two tensors are concatenated along the channel dimension to obtain a tensor of shape l x o. This process not only implements a channel splitting and recombination operation, but also activates the learning of channel information through the inverse residual module and the squeeze-and-excitation layer. Then, the channels are shuffled and activated again, and finally, a post-processing linear layer and a dropout layer are used to output the final vector.

[0064] As an optional implementation, the structure of the frequency domain cross-attention subnetwork is shown in Table 4:

[0065] Table 4 Frequency domain cross-attention subnetwork structure

[0066]

[0067] The frequency domain cross-attention subnetwork receives the output of the first layer of the time domain de-redundancy subnetwork as input, extracts high-frequency components using discrete wavelet transform in the frequency domain gated cross-attention module, and performs cross-attention on the high-frequency components and the original time features. Then, a value of tanh(a) is used as a gate to multiply the result of cross-attention, where a is initialized to 0 and will be optimized with gradient updates. Depthwise separable convolution is used as mapping between layers in this subnetwork. The output of this subnetwork is fed into a selective pooling layer, which outputs a frequency domain feature vector f F . The structure of the frequency domain gated cross-attention module is shown in Table 5:

[0068] Table 5 Frequency domain gated cross-attention module structure

[0069]

[0070] At this time, the frequency domain cross-attention subnetwork and the time domain de-redundancy subnetwork will interact with each other between layers, i.e., the input of the next time domain layer will be added to the output of the previous frequency domain layer, and the input of the next frequency domain layer will also be added to the output of the current time domain layer, forming collaborative feature learning in time and frequency domains.

[0071] As an optional implementation, the context attention feature pyramid sub-network structure is shown in Table 6, the context modeling layer structure is shown in Table 7, and the context modeling module structure is shown in Table 8:

[0072] Table 6: Context attention feature pyramid sub-network structure

[0073]

[0074] Table 7: Context modeling layer structure

[0075]

[0076] Table 8: Context modeling module structure

[0077]

[0078] The context feature pyramid network has two layers, and each layer of the pyramid receives the output of two scales of the layer of the time domain de-redundancy sub-network as input. The two inputs are sent to the context modeling module for context modeling, then one of the features is up-sampled by 2 times, and then added to the feature of the other scale. After addition, a selective pooling module is used to compress the feature vector, and the calculation of one pyramid layer is completed. The time domain features output by the two layers of the pyramid are f T1 and f T2 , respectively. The two features are concatenated in the channel dimension to obtain a feature f T with a shape of 768, which is mapped to a feature with 384 channels through a fully connected layer with an input channel of 768 and an output channel of 384.

[0079] f T is calculated through a fully connected layer to obtain the classification output logit, and the logit calculates the classification loss; f T and f F are input into the metric learning loss to calculate the feature space loss. All losses are added and sent to the optimizer for supervision, and the parameters are updated through gradient backpropagation.

[0080] As an optional implementation, the training parameters of model training are as follows:

[0081] Iteration number: 80;

[0082] Iteration rounds per generation: 2000;

[0083] Optimizer: AdamW;

[0084] Learning rate: 0.001, and the drop strategy is to multiply by 0.9 at the end of each generation;

[0085] Weight decay: 0.00001.

[0086] S4, input the test handwriting and all test handwriting in the database into the trained retrieval model, and extract the feature vectors.

[0087] S5, calculate the cosine similarity between the feature vectors of the test handwriting and the feature vectors of the test handwriting in the database, and obtain the final retrieval result according to the cosine similarity.

[0088] In some optional embodiments, n test handwriting is input into the model to extract the time domain feature vector The frequency domain feature vector is not used. The corresponding time domain feature vector of all test handwriting in the database is extracted. The feature vectors of the test handwriting are traversed, and the cosine similarity between each The cosine similarity is calculated with all the feature vectors in the database, and the order from large to small is sorted, which is the retrieval result. By comparing the retrieval result with the label, the Top1, Top5, Top10 accuracy and mAP can be calculated.

[0089] In summary, the method of the present application has at least the following advantages and beneficial effects compared with the prior art:

[0090] (1) Compared with the previous method for offline writer retrieval, the method proposed in the present application is specifically for online writer retrieval scene, which has a large performance advantage compared with the method migrated from other fields (such as online signature authentication, offline writer retrieval, etc.) to this field, especially in the mAP index. At the same time, the present method is (almost) the first method for online writer retrieval in recent years, and the existing methods are all for offline writer retrieval.

[0091] (2) The present application proposes a channel activation anti-residual module for reducing the redundancy of time domain information channel dimension. The module effectively reduces the redundant information on the channel through multi-level channel activation, such as channel segmentation and recombination, channel expansion and contraction of the anti-residual module, and channel scrambling operation, etc. The module forms a time domain de-redundancy sub-network, which enhances the feature extraction of the model in the time domain.

[0092] (3) The present application proposes a frequency domain gated cross attention module, which first extracts high frequency components using discrete wavelet transform, then performs cross attention on the high frequency components and the original time features, and then uses the value of tanh(a) as the product of the gating and cross attention result. The module forms a frequency domain cross attention sub-network, which excavates the high frequency part of the time domain data as discriminative information from the space, and enhances the learning of the model in the frequency domain.

[0093] (4) The model of the present application interacts the time domain features and the frequency domain features extracted by the two sub-networks with each other between layers, stimulates the common learning effect of the two kinds of features and performs collaborative optimization, promotes the model to better perceive the unique writing style features of the writer in the online handwriting, and achieves better text-independent online writer retrieval performance.

[0094] The present embodiment also provides a text-independent online writer retrieval device, comprising:

[0095] at least one processor;

[0096] at least one memory for storing at least one program;

[0097] When the at least one program is executed by the at least one processor, the at least one processor implements Figure 1 the method shown.

[0098] The text-independent online writer retrieval device of the present embodiment can execute the text-independent online writer retrieval method provided by the method embodiment of the present application, can execute the implementation steps of any combination of the method embodiments, and has the corresponding functions and beneficial effects of the method.

[0099] The present application also discloses a computer program product or a computer program, which comprises computer instructions stored in a computer readable storage medium. The processor of the computer device can read the computer instructions from the computer readable storage medium, and the processor executes the computer instructions to make the computer device execute Figure 1 the method shown.

[0100] The present embodiment also provides a storage medium storing instructions or programs that can execute the text-independent online writer retrieval method provided by the method embodiment of the present application. When the instructions or programs are executed, the implementation steps of any combination of the method embodiments can be executed, and the corresponding functions and beneficial effects of the method are possessed.

[0101] In some alternative embodiments, the functions / operations mentioned in the block diagram can not occur in the order mentioned in the operation diagram. For example, depending on the functions / operations involved, two blocks shown in succession can actually be executed substantially simultaneously or the blocks can sometimes be executed in reverse order. In addition, the embodiments presented and described in the flowcharts of the present application are provided by way of example, with the purpose of providing a more comprehensive understanding of the technology. The disclosed method is not limited to the operations and logical flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and in which sub-operations described as part of larger operations are independently executed.

[0102] Furthermore, although the present application is described in the context of functional modules, it is to be understood that one or more of the described functions and / or features can be integrated in a single physical device and / or software module, or one or more functions and / or features can be implemented in separate physical devices or software modules. It will also be appreciated that detailed discussion of the actual implementation of each module is not necessary to an understanding of the application. Rather, the actual implementation of the modules, in combination with their attributes, functions, and internal relationships, are to be understood within the context of the devices disclosed herein. Thus, those skilled in the art with access to patents, scientific journals, and other public sources known by those skilled in the art will be able, using ordinary skill, to practice the application as set forth in the claims without undue experimentation. It is also to be understood that the specific concepts disclosed are merely illustrative and that the scope of the present application is to be determined by the entire scope of the claims, along with all equivalents of the claims and their equivalents.

[0103] If the functions are implemented in software, the functions can be stored in or implemented as one or more computer program products, which can be incorporated into a computer- readable medium for use by or in connection with an instruction execution system, apparatus, or device, such as a computer-based system, processor- containing system, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions. In the context of this specification, a "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device.

[0104] The logic and / or steps represented in the flowcharts and / or otherwise described herein, for example, can be embodied in non-transitory computer-readable media, which can be executed by an instruction execution system, apparatus, or device such as a computer-based system, processor- containing system, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions. In the context of this specification, a "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device.

[0105] More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection (electronic) having one or more wires, a portable computer diskette (magnetic), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can also be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, for example, via optical scanning of the paper or other medium, then compiled, interpreted, or otherwise processed in a suitable manner, if necessary, and then stored in a computer memory.

[0106] It should be understood that aspects of the application can be implemented in hardware, software, firmware or combinations thereof. In the above described embodiments, various steps or methods can be implemented in software or firmware that is stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any of the following technologies, or combinations thereof, can be used with the necessary modifications: a discrete logic circuit(s) having logic gates for implementing logic functions upon an application of data signals, an application specific integrated circuit having appropriate combinational logic gates, a programmable gate array(s) (PGA), a field programmable gate array (FPGA), and / or the like.

[0107] In the above description of the present specification, reference to the description of the terms "one embodiment", "another embodiment", or "certain embodiments" or the like means that a particular feature, structure, material or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the present application. The illustrative appearances of the above described terms in the specification do not necessarily refer to the same embodiment or example. Moreover, the particular features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0108] Although embodiments of the present application have been shown and described, it would be recognized by those of ordinary skill in the art that various changes, modifications, substitutions, and alterations can be made hereto without departing from the spirit and scope of the application, which is defined by the following claims and their equivalents.

[0109] The above is a specific description of the preferred embodiments of the present application, but the present application is not limited to the above-described embodiments, and those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present application, and these equivalent modifications or substitutions are included in the scope defined by the claims of the present application.

Claims

1. A text-based writer retrieval method, characterized in that, Includes the following steps: Obtain text-independent handwriting time series data; The obtained handwriting time series data is normalized and multiple time functions are extracted to obtain label data; Construct a retrieval model and train it using normalized time-series data and tag data; The test handwriting and all test handwriting samples in the database are input into the trained retrieval model to extract feature vectors; Calculate the cosine similarity between the feature vector of the test handwriting and the feature vector of the test handwriting in the database, and obtain the final retrieval results based on the cosine similarity. The retrieval model is a deep neural network model, which includes a temporal redundancy removal subnetwork, a frequency domain cross-attention subnetwork, and a context-focused feature pyramid network. The temporal redundancy removal subnetwork consists of multiple channel-activated inverse residual modules, which are used to reduce redundancy in the temporal feature channel dimension through multi-level channel activation. The frequency domain cross-attention subnetwork is used to extract the high-frequency components of the time series through discrete wavelet transform, and to perform gated cross-attention calculation on the time series and its high-frequency part to highlight the writing style characteristics of the high-frequency part. The context-focused feature pyramid network is used to obtain the multi-scale output of the temporal redundancy removal sub-network and perform context enhancement, as well as output temporal feature vectors according to the pyramid structure. The frequency domain cross-attention subnetwork and the time domain redundancy removal subnetwork perform feature interaction between layers. That is, the input of the next time domain layer is added to the output of the previous frequency domain layer, and the input of the next frequency domain layer is added to the output of the current time domain layer, forming a collaborative feature learning between the time domain and the frequency domain.

2. The text-based author retrieval method according to claim 1, characterized in that, The normalization process includes pressure normalization and coordinate center normalization.

3. The text-based author retrieval method according to claim 1, characterized in that, The time function includes at least one of the following: the first derivative of the x-coordinate, the first derivative of the y-coordinate, the second derivative of the x-coordinate, the second derivative of the y-coordinate, velocity, the tangent angle between two points, the cosine value of the tangent angle, the sinine value of the tangent angle, the first derivative of the velocity, the first derivative of the tangent angle, the logarithmic radius of curvature, the magnitude of the velocity change, acceleration, or pressure.

4. The text-based author retrieval method according to claim 1, characterized in that, The output of the frequency domain cross-attention subnetwork is pooled to obtain a frequency domain feature vector; When training the model, both time-domain feature vectors and frequency-domain feature vectors are used; when testing the model, only time-domain feature vectors are used for matching between samples.

5. The text-based writer retrieval method according to claim 1, characterized in that, The channel activation anti-residual module operates as follows: The input is split into two tensors along the channel dimension. One of these tensors is processed by an anti-residual module. Dimension mapping to The dimension is then passed through a squeezing and activation layer to activate the channels, and finally mapped to a convolutional-nonlinear layer. Dimension; another tensor is directly passed from a convolutional layer used for mapping. Mapped to Dimension; concatenating these two tensors along the channel dimension yields a shape of... The tensor; where, For the input channel dimension, For sequence length, For intermediate layer channel dimensions, Output channel dimension; The channels are shuffled and activated again. Finally, the vector is output through a post-processing linear layer and a dropout layer.

6. The text-based writer retrieval method according to claim 1, characterized in that, The frequency domain cross-attention subnetwork operates as follows: The frequency-domain cross-attention subnetwork receives the output of the first layer of the time-domain redundancy removal subnetwork as input, extracts high-frequency components using discrete wavelet transform, then performs cross-attention on the high-frequency components and the original time features, and finally uses a tanh(a) value as a gate to multiply the result of the cross-attention. The output of this frequency-domain cross-attention subnetwork is fed into a selective pooling layer, outputting a frequency-domain feature vector. .

7. The text-based writer retrieval method according to claim 1, characterized in that, The context-focused feature pyramid network works as follows: The context-focused feature pyramid network consists of two pyramid layers. Each pyramid layer receives the outputs of the two-scale temporal redundancy-removal subnetwork layers as inputs. The two inputs are fed into the context building module for context modeling. Then, one feature is upsampled and added to the feature at the other scale. After addition, it is compressed into a feature vector by a selective pooling module to complete the calculation of one pyramid layer. The temporal characteristics of the two-layer pyramid output are as follows: and The two time-domain features are concatenated along the channel dimension to obtain the time-domain features. .

8. A text-based writer retrieval device, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method of any one of claims 1-7.

9. A computer-readable storage medium storing a processor-executable program, characterized in that, The processor-executable program, when executed by the processor, is used to perform the method as described in any one of claims 1-7.

Images