Data feature extraction method and device, equipment and medium

Abstract

The invention relates to the technical field of image processing, can be applied to business system platforms of financial science and technology, medical health and the like, and discloses a data feature extraction method, device, equipment and medium, the method comprises the following steps: obtaining an initial to-be-detected image, and extracting an initial feature sequence of the initial to-be-detected image, performing context dependency capture and coding on the initial feature sequence to obtain a plurality of coding sequence features, determining an importance score of each coding sequence feature, and dynamically screening the importance scores according to a preset adaptive threshold to obtain key sequence features corresponding to the screened importance scores, and performing pyramid fusion on the key sequence features to obtain fused sequence features, and performing feature recognition on the initial to-be-detected image according to the fused sequence features to obtain target data features. According to the invention, the accuracy and efficiency of data feature extraction can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a data feature extraction method, device, equipment and medium. Background Art

[0002] In insurance data processing, accurately identifying and extracting key features (such as vehicle damage locations and abnormal indicators in medical records) is crucial for damage assessment and claims settlement decisions. However, in real-world applications, insurance data often faces multiple challenges, including partial occlusion, noise interference, inconsistent data formats, and strong temporal dependencies. These complex factors result in poor accuracy and efficiency in traditional information extraction methods.

[0003] In the healthcare sector, the medical images, medical records, and examination reports that claims rely on usually contain a large amount of unstructured information, and the data quality varies. For example, medical record descriptions are often interspersed with subjective expressions, inconsistent terminology, or non-standard writing. In addition, medical images are blurry, occluded, or have low resolution, making it difficult to accurately extract and quantify key indicators (such as abnormal values ​​and diagnostic conclusions). Traditional methods rely on rules or template processing, and often make misjudgments or omissions in the face of complex and changing inputs, resulting in low recognition accuracy. In addition, because the data parsing process relies heavily on manual review and the processing cycle is long, the overall efficiency is difficult to meet the needs of claims automation, which directly affects the timeliness of claims and customer experience.

[0004] In the fintech sector, insurance data processing often involves highly time-sensitive data types such as user behavior records, transaction flows, and risk assessment results. However, this data often comes from multiple sources, has inconsistent formats, and is subject to omissions, duplications, and delays, severely limiting the accuracy of information fusion and time-series modeling. Accurately extracting key trigger features is crucial in automated underwriting and anti-fraud scenarios. Traditional statistical or rule-based engines struggle to cope with complex semantics and dynamic changes, resulting in limited recognition effectiveness and high error rates. Furthermore, large-scale data analysis relies on high-performance computing resources, and traditional processing methods are inefficient in meeting the requirements of real-time response.

[0005] In summary, problems such as the complexity of unstructured information, inconsistent quality of multi-source data, and strong time dependence cause traditional methods to often make misjudgments and omissions when extracting key features, making it difficult to meet the needs of automated processing and real-time decision-making.

[0006] Therefore, current technologies have the problems of low accuracy and low efficiency in data feature extraction. Summary of the Invention

[0007] The present invention provides a data feature extraction method, device, equipment and medium, the main purpose of which is to solve the problems of low accuracy and low efficiency in data feature extraction.

[0008] In a first aspect, to achieve the above-mentioned purpose, the present invention provides a data feature extraction method, comprising:

[0009] Acquire an initial image to be detected, and extract an initial feature sequence of the initial image to be detected;

[0010] Capturing and encoding context dependencies of the initial feature sequence to obtain a plurality of coded sequence features;

[0011] Determining the importance score of each of the coding sequence features, and dynamically screening the importance scores according to a preset adaptive threshold to obtain key sequence features corresponding to the screened importance scores;

[0012] Performing pyramid fusion on the key sequence features to obtain fused sequence features;

[0013] Feature recognition is performed on the initial image to be detected according to the fusion sequence features to obtain target data features.

[0014] In a second aspect, the present invention further provides a data feature extraction device, comprising:

[0015] A feature sequence extraction module is used to obtain an initial image to be detected and extract an initial feature sequence of the initial image to be detected;

[0016] A dependency capturing module, configured to capture and encode context dependencies of the initial feature sequence to obtain a plurality of encoding sequence features;

[0017] A key feature screening module is used to determine the importance score of each coding sequence feature and dynamically screen the importance score according to a preset adaptive threshold to obtain the key sequence feature corresponding to the screened importance score;

[0018] A key feature fusion module is used to perform pyramid fusion on the key sequence features to obtain fused sequence features;

[0019] The feature recognition module is used to perform feature recognition on the initial image to be detected based on the fusion sequence features to obtain target data features.

[0020] In a third aspect, the present invention further provides an electronic device, comprising:

[0021] at least one processor; and,

[0022] a memory communicatively connected to the at least one processor; wherein,

[0023] The memory stores a computer program that can be executed by the at least one processor. The computer program is executed by the at least one processor so that the at least one processor can perform the data feature extraction method described above.

[0024] In a fourth aspect, the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores at least one computer program, and the at least one computer program is executed by a processor in an electronic device to implement the above-mentioned data feature extraction method.

[0025] The present invention obtains an initial image to be detected and extracts an initial feature sequence of the initial image to be detected. Multi-layer convolutional layers can capture rich structural and texture information in the image at multiple scales and levels. Average pooling effectively reduces feature dimensions, enhances the translation invariance of features, and improves computational efficiency. The activation function introduces nonlinear transformations to enhance feature expression capabilities and highlight key areas. The context dependency of the initial feature sequence is captured and encoded to obtain a number of coding sequence features. The bidirectional capture model is used to perform forward and reverse recursive calculations and aggregation on the initial feature sequence to fully explore the context dependency between adjacent and distant elements in the sequence, thereby improving the semantic expression capabilities of the features. The importance score of each coding sequence feature is determined, and the importance score is dynamically screened according to a preset adaptive threshold to obtain key sequence features corresponding to the screened importance scores, and weights are assigned to the coding sequence features. The vectors are vectored and the dot product score is calculated to quantify the importance of each feature. The scores are ensured to be comparable after normalization mapping. A screening strategy based on a preset adaptive threshold and a dynamic weighted smoothing mechanism is adopted to effectively balance the influence of historical thresholds and current scores, realize dynamic identification and filtering of key sequence features, perform pyramid fusion on the key sequence features, obtain fused sequence features, use bilinear interpolation to spatially align the convolution features, ensure the consistency of features of different scales in the spatial dimension, and fuse the multi-scale aligned features through splicing operations to enrich the diversity and integrity of feature expression. Feature recognition is performed on the initial image to be detected according to the fused sequence features to obtain target data features, which are fused by combining the discriminant features with the confidence level to generate target data features, which not only reflects the key information of the image, but also includes the confidence level of the results, which can effectively improve the accuracy and efficiency of data feature extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments of the present invention. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0027] Figure 1 A schematic diagram of an application environment of a data feature extraction method according to an embodiment of the present invention;

[0028] Figure 2 A flow chart of a data feature extraction method provided by one embodiment of the present invention;

[0029] Figure 3 A schematic flow chart of a pyramid fusion process in a data feature extraction method provided by one embodiment of the present invention;

[0030] Figure 4 A schematic diagram of a module of a data feature extraction device provided by one embodiment of the present invention;

[0031] Figure 5 A schematic structural diagram of an electronic device for implementing a data feature extraction method provided by one embodiment of the present invention;

[0032] Figure 6 Another structural diagram of an electronic device for implementing a data feature extraction method provided by an embodiment of the present invention.

[0033] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0034] In order to enable those skilled in the art to better understand the technical solutions of the present disclosure, and to fully understand and implement how the present disclosure applies technical means to solve technical problems and achieve the corresponding technical effects, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all of the embodiments. The embodiments of the present disclosure and the various features in the embodiments can be combined with each other without conflict, and the technical solutions formed are all within the scope of protection of the present disclosure. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without making creative work should fall within the scope of protection of the present disclosure.

[0035] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, apparatus, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0036] An embodiment of the present application provides a data feature extraction method, and the execution subject of the data feature extraction method includes but is not limited to at least one of the electronic devices such as a server, a terminal, etc. that can be configured to execute the device provided by the embodiment of the present application. In other words, the data feature extraction method can be executed by software or hardware installed in a terminal device or a server device. The server includes but is not limited to: a single server, a server cluster, a cloud server or a cloud server cluster, etc. The server can be an independent server, or it can be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0037] The present invention provides a data feature extraction method, which can be applied to Figure 1application environment. Among them, the client communicates with the server through the network. The server can obtain the initial image to be detected through the client, extract the initial feature sequence of the initial image to be detected, and the multi-layer convolution layer can capture the rich structure and texture information in the image at multiple scales and levels. The average pooling effectively reduces the feature dimension, enhances the translation invariance of the feature, and improves the computational efficiency. The activation function introduces nonlinear transformation, enhances the feature expression ability, and highlights the key areas. The context dependency relationship of the initial feature sequence is captured and encoded to obtain a number of coding sequence features. The bidirectional capture model is used to perform forward and reverse recursive calculations and aggregation on the initial feature sequence, fully exploring the context dependency relationship between adjacent and distant elements in the sequence, and improving the semantic expression ability of the feature. The importance score of each coding sequence feature is determined, and the importance score is dynamically screened according to a preset adaptive threshold to obtain the key sequence features corresponding to the screened importance score. The coding sequence features are given a weight vector and the dot product is calculated. The importance of each feature is quantified, and normalized mapping is performed to ensure comparability of scores. A screening strategy based on a preset adaptive threshold and a dynamic weighted smoothing mechanism is adopted to effectively balance the influence of historical thresholds and current scores, achieving dynamic identification and filtering of key sequence features. The key sequence features are pyramidally fused to obtain fused sequence features. The convolution features are spatially aligned using bilinear interpolation to ensure consistency of features at different scales in the spatial dimension. The multi-scale aligned features are fused through a splicing operation to enrich the diversity and integrity of feature expression. Feature recognition is performed on the initial image to be detected based on the fused sequence features to obtain target data features. The discriminant features are combined with the confidence level to generate target data features. The target data features reflect the key information of the image and also include the confidence level of the results, which can effectively improve the accuracy and efficiency of data feature extraction. Finally, the target data features are output and fed back to the user client. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented as an independent server or a server cluster consisting of multiple servers. The present invention is described in detail below through specific embodiments.

[0038] The following is an explanation of the description of the present invention. The present invention effectively balances the influence of historical thresholds and current scores by adopting a screening strategy based on preset adaptive thresholds and a dynamic weighted smoothing mechanism, thereby realizing dynamic identification and filtering of key sequence features. This not only improves the flexibility and robustness of feature selection, avoids the limitations brought by fixed thresholds, but also enhances the model's ability to focus on important information, suppresses redundant and noise features, and thus significantly improves the accuracy and efficiency of data feature extraction.

[0039] Reference Figure 2FIG. 1 is a flow chart of a data feature extraction method provided by an embodiment of the present invention. In this embodiment, the data feature extraction method includes:

[0040] S1. Acquire an initial image to be detected, and extract an initial feature sequence of the initial image to be detected.

[0041] In an embodiment of the present invention, an initial image to be detected, such as a vehicle damage assessment image, time series data of medical or claims records, etc., is input into a pre-trained lightweight convolutional neural network (such as MobileNetV3). Local visual features of the initial image to be detected are extracted through multi-layer convolution and activation operations, and the extracted two-dimensional feature map is flattened and serialized into a set of feature vectors with consistent dimensions according to spatial position to form an initial feature sequence, which provides basic input for subsequent time series modeling and attention mechanism.

[0042] In specific healthcare scenarios, such as intelligent analysis of electronic medical records, such as processing insurance claims that include brain MRI images, the MRI image is first acquired and its initial feature sequence is extracted. This provides a basis for subsequent identification of lesion areas or tissue abnormalities (such as bleeding points and tumor boundaries), supports automatic assessment of disease severity and claim rationality, and reduces the time and subjective errors of manual interpretation by doctors.

[0043] In specific FinTech scenarios, this technology is used, for example, to process identity verification images or transaction credential images (such as occluded ID photos or scanned receipts) in user risk assessments. Extracting the initial feature sequence after acquiring the uploaded image is used to subsequently identify key fields or verify image authenticity, helping to improve anti-fraud systems' ability to identify forgery and tampering, ensuring the compliance and security of financial transactions.

[0044] In an embodiment of the present invention, extracting the initial feature sequence of the initial image to be detected includes:

[0045] Performing denoising processing on the initial image to be detected to obtain a denoised image;

[0046] Convolving the denoised image to obtain a convolved image;

[0047] Performing average pooling on the convolution image to obtain a pooled image;

[0048] Activating the pooled image using a preset activation function to obtain spatial local features;

[0049] A spatial dimension is obtained, and the spatial local features are mapped into an initial feature sequence according to the spatial dimension.

[0050] In detail, according to the type of image noise, a suitable denoising algorithm is selected, such as median filtering, bilateral filtering or wavelet transform, and the image is processed pixel by pixel or block by block to suppress interference information such as random noise, salt and pepper noise or Gaussian noise. Finally, a denoised image with good texture details and clear edges is obtained, providing a more reliable input basis for subsequent image analysis or recognition.

[0051] The denoised image is input into the convolutional neural network, and its feature information at different levels is extracted through multi-layer convolution operations. Each layer of convolution uses several convolution kernels to perform sliding calculations on the image to capture local edges, textures, shapes and other features. The convolution result is processed by a nonlinear activation function and passed to the next layer, thereby gradually constructing a deeper and more abstract image representation, and finally obtaining a convolution image containing rich feature information.

[0052] The convolution image is divided into several non-overlapping or partially overlapping regions, and the average value of all pixel values ​​in each region is calculated as the representation value of the region in the pooled image. This can effectively reduce the image size and feature dimension while retaining the main image information.

[0053] Each pixel value in the pooled image is input into a preset nonlinear activation function (such as ReLU, Sigmoid or Tanh), and the pixel value is mapped point by point. The activation operation can enhance the expression ability of nonlinear features in the image, suppress invalid or negative responses, and highlight the response intensity of key areas. Through the activation process, spatial local features with discriminative significance can be effectively extracted.

[0054] The spatial dimension information such as the height, width and number of channels of the local spatial features is extracted, and the three-dimensional feature matrix is ​​flattened according to the preset arrangement rules (such as row priority or column priority). The multi-channel feature vectors at each spatial position are expanded and connected in sequence to convert them into a one-dimensional feature vector sequence.

[0055] By first denoising, the image quality is improved, noise interference is reduced, and the accuracy of subsequent feature extraction is guaranteed; multi-layer convolutional layers can capture rich structural and texture information in the image at multiple scales and levels; average pooling effectively reduces feature dimensions, enhances the translation invariance of features, and improves computational efficiency; the activation function introduces nonlinear transformations, enhances feature expression capabilities, and highlights key areas; finally, through spatial dimension mapping, the two-dimensional features are converted into sequence form, which facilitates subsequent sequence modeling and processing, so that the extracted initial feature sequence has both rich semantic information and good structured feature expression, laying a solid foundation for subsequent image analysis tasks.

[0056] S2. Capturing and encoding context dependencies of the initial feature sequence to obtain a plurality of encoding sequence features.

[0057] In this embodiment of the present invention, the initial feature sequence is input into a bidirectional gated recurrent unit (BiGRU), where feature modeling is performed along both the forward and reverse time steps to capture the contextual dependencies between positions in the image sequence. This encoding process effectively integrates local spatial features with global temporal information, generating a set of context-aware encoded sequence features. This provides a semantically rich input foundation for subsequent important region identification and feature enhancement.

[0058] In specific medical and health scenarios, for example, it is used to automatically interpret continuous-slice medical images (such as CT or MRI sequences). By modeling and encoding the correlation features between slices in the image, it can accurately capture the development trend of lesions or the boundaries of lesions, thereby assisting doctors in identifying potential high-risk areas and improving the accuracy of judgment on the disease evolution process in claims.

[0059] In specific financial technology scenarios, for example, it is used to process user behavior logs or transaction image sequences, such as analyzing a series of bills or identity verification images uploaded in chronological order. By judging the contextual dependencies between encoded images, the integrity and consistency between encoded images can be determined, thereby identifying whether there are fraudulent behaviors such as repeated submissions and information forgery, thereby improving the intelligent level of risk control.

[0060] In the embodiment of the present invention, the context dependency relationship of the initial feature sequence is captured and encoded to obtain a plurality of encoding sequence features, including:

[0061] Recursively calculating and summarizing the initial feature sequence in a preset first direction and a preset second direction using a preset bidirectional capture model to obtain a hidden state sequence group;

[0062] splicing the sequences in the hidden state sequence group to obtain context-enhanced features;

[0063] Non-linear activation is performed on the context enhancement feature to obtain a coding sequence feature.

[0064] In detail, the initial feature sequence is input into the forward and reverse sub-networks of a bidirectional recurrent neural network model (such as Bi-LSTM or Bi-GRU), respectively. In the forward network, each feature vector is processed sequentially from the start of the sequence, while in the reverse network, it is processed in reverse order from the end of the sequence. The hidden state at each moment is determined by the current input and the state at the previous moment, gradually capturing context-dependent information, and finally obtaining two-directional hidden state sequence groups, which respectively characterize the information flow of the sequence in the forward and reverse directions.

[0065] The forward hidden state vector and the reverse hidden state vector corresponding to each time step in the sequence are concatenated element by element in the feature dimension to form a new higher-dimensional vector. In this way, the concatenated context-enhanced feature contains both historical information from the beginning of the sequence to the current position and future information from the end of the sequence to the current position, thus achieving bidirectional integration of the context. The calculation formula is as follows:

[0066]

[0067] in, represents the hidden state sequence of the t-th time step in the preset first direction, GRU represents the bidirectional capture model of the recurrent neural network, x t represents the initial feature sequence at the t-th time step, represents the hidden state sequence at the t-1th time step in the preset first direction, represents the hidden state sequence at the t-th time step in the preset second direction, represents the hidden state sequence at the t+1th time step in the preset second direction, h t Represents context-enhanced features.

[0068] When processing context-enhanced features, the GELU (Gaussian Error Linear Unit) activation function is introduced to perform nonlinear transformations on the context-enhanced features. Each context feature vector encoded by the bidirectional GRU is input into the GELU function, which combines the input value with the cumulative distribution characteristics of the standard Gaussian distribution to achieve smooth activation of the feature. Compared to the traditional ReLU activation method, GELU can more naturally retain the fine-grained information of the input, improve sensitivity to feature changes, and help improve the model's expressiveness and training stability when processing complex insurance images or text data. The calculation formula is as follows:

[0069] H=GELU(h t )=h t Φ(h t )

[0070] Among them, GELU represents the nonlinear activation function, h t represents the context-enhanced feature, Φ(h t ) represents the Gaussian error function, and H represents the encoding sequence feature.

[0071] By utilizing the bidirectional capture model to perform forward and backward recursive calculations and summaries on the initial feature sequence, the contextual dependencies between adjacent and distant elements in the sequence are fully exploited, thereby improving the semantic expression capability of the features. The concatenation of hidden state sequences integrates bidirectional information, enriches the temporal and spatial semantic dimensions of the features, and makes the context-enhanced features more comprehensive and more discriminative. The activation operation further enhances the nonlinear expression capability and strengthens the response of key features. The resulting coded sequence features have a deeper level of semantic understanding and structural information, which contributes to the efficient completion of subsequent precise recognition and classification tasks.

[0072] S3. Determine the importance score of each coding sequence feature, and dynamically screen the importance scores according to a preset adaptive threshold to obtain key sequence features corresponding to the screened importance scores.

[0073] In an embodiment of the present invention, an importance score corresponding to each encoded sequence feature is calculated to measure the contribution of the sequence feature to the overall recognition task. An adaptive threshold mechanism based on sliding average is introduced to dynamically screen these importance scores, and only key sequence features with scores higher than the current threshold are retained. The screening criteria can be automatically adjusted according to the specific characteristics of the input data, thereby highlighting key information that is of practical significance to the final judgment, while suppressing redundancy or noise interference, thereby improving the accuracy and efficiency of subsequent recognition.

[0074] In specific medical and health scenarios, it can be used to automatically screen out key areas related to lesions from a series of medical images. For example, in lung CT scans, it can dynamically identify nodule areas with significant density changes and exclude irrelevant or normal tissue images, thereby improving the accuracy of abnormality detection.

[0075] In specific FinTech scenarios, this technology can be applied to analyzing user transaction behavior sequences, such as automatically identifying high-risk segments in user transactions within risk control systems, such as unusual time periods, frequent changes in location, or device usage history. By dynamically screening the most critical behavioral features for risk assessment, it helps improve the accuracy of fraud detection and reduce false positives.

[0076] In an embodiment of the present invention, determining the importance score of each coding sequence feature includes:

[0077] Obtaining a weight vector for each of the coding sequence features, performing a dot product between each of the coding sequence features and the corresponding weight vector to obtain a dot product score;

[0078] The point product scores are normalized and mapped to obtain importance scores.

[0079] In detail, the calculation formula of the importance score is as follows:

[0080] α t =σ(W α H+b α )

[0081] Among them, W α represents the weight vector, σ represents the normalization factor, H represents the encoding sequence feature, b α represents the learning bias parameter in the importance score calculation process, α t Represents the importance score.

[0082] In an embodiment of the present invention, the importance scores are dynamically screened according to a preset adaptive threshold to obtain key sequence features corresponding to the screened importance scores, including:

[0083] Obtaining an initial threshold value of a preset time step and an importance score of a next time step of the preset time step;

[0084] Obtaining a smoothing coefficient, and generating a weighted weight according to the smoothing coefficient;

[0085] Performing a weighted average of the initial threshold value and the importance score of the next time step using the weighted weight to obtain an updated threshold value;

[0086] Determining whether the importance score of the next time step is greater than the threshold update value;

[0087] If the importance score of the next time step is less than or equal to the threshold update value, deleting the coding sequence feature corresponding to the importance score of the next time step;

[0088] If the importance score of the next time step is greater than the threshold update value, filtering out the coding sequence feature corresponding to the importance score of the next time step, and confirming the activation attention of the filtered coding sequence feature;

[0089] Key sequence features are generated based on the activated attention and the filtered coding sequence features.

[0090] In detail, the initial value of the threshold corresponding to the preset time step and the importance score of the next time step are obtained, and the corresponding weighted weight is calculated according to the preset smoothing coefficient to balance the influence ratio of the historical threshold and the current score in the threshold update. Then, the weighted weight is used to multiply the initial value of the threshold and the importance score of the next time step respectively, and the results are added to complete the weighted average calculation, and finally the updated threshold is obtained.

[0091] By comparing the importance score of the next time step with the dynamically updated threshold, key and non-key features can be effectively distinguished: when the score is lower than or equal to the threshold, the corresponding coding sequence feature is automatically shielded to prevent redundant or noise information from interfering with the model; when the score is higher than the threshold, the feature is promptly confirmed as a key sequence feature to ensure the retention and utilization of important information. This dynamic and adaptive screening mechanism improves the accuracy and efficiency of feature selection.

[0092]

[0093] Among them, τ represents the threshold update value, T represents the total time, t represents the preset time step, α t represents the importance score, represents the average importance score, and λ represents a hyperparameter.

[0094] The system confirms the activated attention mechanism of the screened coding sequence features, indicating that the model has assigned significant attention weights to the screened coding sequence feature positions. The activated attention of the screened coding sequence features is fused with the corresponding coding features, such as through weighted operations or gated combinations, to generate key sequence features with contextual weight information. The activated attention calculation formula is as follows:

[0095]

[0096] Among them, softmax represents the normalization function, represents the activation parameter matrix, tanh represents the hyperbolic tangent function, W β represents weighted weight, H represents the coding sequence feature, b β represents the learning bias parameter in the activation attention calculation process, β t Indicates activation of attention.

[0097] By assigning weight vectors to the coding sequence features and calculating the dot product scores, the importance of each feature is quantified, and the scores are ensured to be comparable after normalization mapping; a screening strategy based on preset adaptive thresholds and dynamic weighted smoothing mechanism is adopted to effectively balance the influence of historical thresholds and current scores, and realize dynamic identification and filtering of key sequence features. This not only improves the flexibility and robustness of feature selection, avoids the limitations brought by fixed thresholds, but also strengthens the model's ability to focus on important information, suppresses redundant and noisy features, and thus significantly improves the accuracy and efficiency of subsequent tasks.

[0098] S4. Perform pyramid fusion on the key sequence features to obtain fused sequence features.

[0099] In this embodiment of the present invention, the selected key sequence features are input into a pyramid fusion module, which extracts feature information from different levels of representation. Upsampling aligns the spatial dimensions of the features at each level, integrating multi-scale information. During the fusion process, normalization operations (such as LayerNorm) are introduced to stabilize training and suppress feature bias, thereby generating fused sequence features with richer semantic layers and stronger discriminative power.

[0100] In specific medical and health scenarios, for example, it is used to fuse medical image features at different depths and resolutions. For example, in tumor detection, coarse-grained organ contour information and fine-grained lesion edge features are simultaneously integrated to enhance the recognition ability of lesion areas from multiple scales.

[0101] In specific FinTech scenarios, this can be used to integrate risk profiles across different time periods or dimensions, such as combining past user behavior patterns, current transaction behavior, and device environment characteristics to form a multi-layered risk profile. This multi-scale integration allows for a more comprehensive assessment of the potential risks behind user behavior.

[0102] Figure 3 A flowchart of a pyramid fusion process in a data feature extraction method provided by one embodiment of the present invention is shown.

[0103] In an embodiment of the present invention, performing pyramid fusion on the key sequence features to obtain fused sequence features includes:

[0104] Performing multi-layer convolution on the key sequence features to obtain several key convolution features;

[0105] Performing bilinear interpolation on the key convolution features to obtain aligned convolution features;

[0106] Splicing the aligned convolution features to obtain spliced ​​convolution features;

[0107] A scaling parameter and a translation parameter are obtained, and the concatenated convolution features are normalized using the scaling parameter and the translation parameter to obtain a fused sequence feature.

[0108] In detail, the encoded key sequence features are input into a multi-layer one-dimensional or two-dimensional convolutional network. Each layer of convolution extracts features by sliding multiple convolution kernels in the local area, gradually capturing structural patterns and semantic information of different scales and levels. The convolution results are processed by the activation function and input into the next layer to deepen the network's understanding and abstract expression of key features, and finally obtain several key convolution features. These features strengthen global correlation while retaining local details.

[0109] Bilinear interpolation is performed on the key convolution feature map in the spatial dimension. The new pixel value is smoothly estimated by taking the weighted average of the four adjacent pixels in the key convolution feature map at each interpolation position, thereby achieving continuous scaling of the feature map in the horizontal and vertical directions, effectively adjusting the spatial size of the feature map, realizing multi-scale feature alignment, and generating aligned convolution features with consistent size and continuous content while maintaining structural information and edge smoothness.

[0110] Multiple aligned convolutional features are connected point by point in the channel dimension to form a higher-dimensional feature representation. Through this splicing operation, features from different sources or different scales can be fused, enriching the expressive power of the features, and finally obtaining the spliced ​​convolutional features.

[0111] After obtaining the scaling parameter and translation parameter, we use these two parameters to perform linear transformation on each channel of the concatenated convolution feature element by element, and adjust the feature value to a unified numerical range, which helps to alleviate the differences in different feature scales and distributions, improve the stability and convergence speed of model training, and finally obtain the fusion sequence feature. Let the output of the first layer BiGRU be C (l) , the fusion process is:

[0112] F=LayerNorm(C (1) +U2C (2) +U3C (3) )

[0113] Among them, C (1) Represents the first layer of key convolution features, C (2) Represents the second layer key convolutional features, C (3) Represents the third-layer key convolutional features, U2 represents the second-layer aligned convolutional features, U3 represents the third-layer aligned convolutional features, LayerNorm represents layer normalization, and F represents fused sequence features.

[0114] Through multi-layer convolution, multi-scale and multi-level feature extraction of key sequence features is performed to capture rich detail information at different scales in the image; bilinear interpolation is used to spatially align the convolution features to ensure the consistency of features of different scales in the spatial dimension; multi-scale aligned features are fused through splicing operations to enrich the diversity and completeness of feature expression; the spliced ​​features are normalized by combining scaling and translation parameters to effectively eliminate scale and displacement differences, achieve unified and standardized feature representation, improve the expressive power and stability of the fused sequence features, and provide a more accurate and robust feature foundation for subsequent recognition, classification or detection tasks.

[0115] S5. Perform feature recognition on the initial image to be detected according to the fusion sequence features to obtain target data features.

[0116] In this embodiment of the present invention, the fused sequence features are input into a task-specific prediction module for final feature recognition of the initial image to be inspected. By interpreting the multi-scale, contextual, and key region information contained in the fused features, the system accurately identifies target data features in the image, such as injury type, lesion classification, or risk label. This outputs a structured recognition result, providing a direct basis for subsequent damage assessment, claims settlement, or risk assessment.

[0117] In specific medical and health scenarios, for example, it is used to identify and classify lesions in medical images based on fusion features, such as accurately locating and determining the type of tumor and its severity, helping doctors quickly generate diagnostic reports and claims recommendations, improving diagnostic efficiency and accuracy, and reducing the workload of medical staff.

[0118] In specific financial technology scenarios, for example, it is used to combine fusion features to identify and classify risks in transaction credentials or identity verification images submitted by users, accurately judge the legality of transactions and potential fraud risks, and thus assist the risk control system in making real-time and accurate risk warnings.

[0119] In an embodiment of the present invention, performing feature recognition on the initial image to be detected based on the fusion sequence features to obtain target data features includes:

[0120] Acquire a target analysis task, map the fused sequence features to a target space according to the target analysis task, and extract discriminant features of the initial image to be detected in the target space;

[0121] Determining a confidence value of the target analysis task based on the discriminant feature;

[0122] A target data feature is generated according to the discriminant feature and the confidence value.

[0123] In detail, after obtaining the target analysis task, the fused sequence features are converted to the target space through a mapping function (such as a fully connected layer, a projection layer, or a task-specific transformation module) according to the specific requirements of the task. In the target space, the discriminant features related to the initial image to be detected are extracted. These features can effectively reflect the key attributes or category information of the image, realize task-driven focusing and screening of the fused features, and ensure that the discriminant features are highly distinctive and discriminative.

[0124] The discriminant features are input into the confidence assessment module (such as the soft max layer, probability estimator or regression model). By calculating the degree of matching between the discriminant features and each category, the corresponding confidence value is generated, which reflects the reliability and certainty of the target analysis task results. The higher the value, the more confident the model is in the current judgment.

[0125] The discriminant features are fused with the corresponding confidence values, usually through weighting, splicing or fusion networks, and the confidence values ​​are used as weights to guide the enhancement or adjustment of the discriminant features, effectively combining the discriminant information of the features with the credibility of the results, improving the expression accuracy and robustness of the target data features, and obtaining the target data features.

[0126] By combining fusion sequence features and specific target analysis tasks, the rich features of multi-scale fusion are mapped to the target space, and the discriminant features related to the initial image to be detected are accurately extracted to achieve highly targeted feature recognition; the confidence value is calculated using the discriminant features to effectively quantify the reliability of the recognition results and enhance the credibility of the model judgment; the discriminant features are combined with the confidence to generate the target data features, which not only reflects the key information of the image, but also includes the confidence level of the results, which not only improves the accuracy and efficiency of data feature extraction, but also enhances the model's adaptability to complex tasks and the reliability of decision-making.

[0127] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0128] like Figure 4 , which is a functional module diagram of a data feature extraction device provided by one embodiment of the present invention.

[0129] In an embodiment of the present disclosure, a data feature extraction device is provided, and the data feature extraction device corresponds one-to-one to the data feature extraction method in the above embodiment. Figure 4 As shown, the data feature extraction device 100 can be installed in an electronic device. According to the functions to be implemented, the data feature extraction device 100 includes a feature sequence extraction module 101, a dependency capture module 102, a key feature screening module 103, a key feature fusion module 104, and a feature recognition module 105. The functional modules are described in detail as follows:

[0130] The feature sequence extraction module 101 is used to obtain an initial image to be detected and extract an initial feature sequence of the initial image to be detected;

[0131] A dependency capturing module 102 is configured to capture and encode context dependencies of the initial feature sequence to obtain a plurality of encoded sequence features;

[0132] The key feature screening module 103 is used to determine the importance score of each coding sequence feature and dynamically screen the importance score according to a preset adaptive threshold to obtain the key sequence feature corresponding to the screened importance score;

[0133] A key feature fusion module 104 is used to perform pyramid fusion on the key sequence features to obtain fused sequence features;

[0134] The feature recognition module 105 is configured to perform feature recognition on the initial image to be detected based on the fusion sequence features to obtain target data features.

[0135] In one embodiment, when extracting the initial feature sequence of the initial image to be detected, the feature sequence extraction module 101 is configured to:

[0136] Performing denoising processing on the initial image to be detected to obtain a denoised image;

[0137] Convolving the denoised image to obtain a convolved image;

[0138] Performing average pooling on the convolution image to obtain a pooled image;

[0139] Activating the pooled image using a preset activation function to obtain spatial local features;

[0140] A spatial dimension is obtained, and the spatial local features are mapped into an initial feature sequence according to the spatial dimension.

[0141] In one embodiment, when the dependency capturing module 102 performs context dependency capturing and encoding on the initial feature sequence to obtain a plurality of encoded sequence features, it is configured to:

[0142] Recursively calculating and summarizing the initial feature sequence in a preset first direction and a preset second direction using a preset bidirectional capture model to obtain a hidden state sequence group;

[0143] splicing the sequences in the hidden state sequence group to obtain context-enhanced features;

[0144] Non-linear activation is performed on the context enhancement feature to obtain a coding sequence feature.

[0145] In one embodiment, when determining the importance score of each coding sequence feature, the key feature screening module 103 is configured to:

[0146] Obtaining a weight vector for each of the coding sequence features, performing a dot product between each of the coding sequence features and the corresponding weight vector to obtain a dot product score;

[0147] The point product scores are normalized and mapped to obtain importance scores.

[0148] In one embodiment, when the key feature screening module 103 dynamically screens the importance scores according to a preset adaptive threshold to obtain key sequence features corresponding to the screened importance scores, it is configured to:

[0149] Obtaining an initial threshold value of a preset time step and an importance score of a next time step of the preset time step;

[0150] Obtaining a smoothing coefficient, and generating a weighted weight according to the smoothing coefficient;

[0151] Performing a weighted average of the initial threshold value and the importance score of the next time step using the weighted weight to obtain an updated threshold value;

[0152] Determining whether the importance score of the next time step is greater than the threshold update value;

[0153] If the importance score of the next time step is less than or equal to the threshold update value, deleting the coding sequence feature corresponding to the importance score of the next time step;

[0154] If the importance score of the next time step is greater than the threshold update value, filtering out the coding sequence feature corresponding to the importance score of the next time step, and confirming the activation attention of the filtered coding sequence feature;

[0155] Key sequence features are generated based on the activated attention and the filtered coding sequence features.

[0156] In one embodiment, when performing pyramid fusion on the key sequence features to obtain fused sequence features, the key feature fusion module 104 is configured to:

[0157] Performing multi-layer convolution on the key sequence features to obtain several key convolution features;

[0158] Performing bilinear interpolation on the key convolution features to obtain aligned convolution features;

[0159] Splicing the aligned convolution features to obtain spliced ​​convolution features;

[0160] A scaling parameter and a translation parameter are obtained, and the concatenated convolution features are normalized using the scaling parameter and the translation parameter to obtain a fused sequence feature.

[0161] In one embodiment, when the feature recognition module 105 performs feature recognition on the initial image to be detected based on the fusion sequence features to obtain target data features, it is configured to:

[0162] Acquire a target analysis task, map the fused sequence features to a target space according to the target analysis task, and extract discriminant features of the initial image to be detected in the target space;

[0163] Determining a confidence value of the target analysis task based on the discriminant feature;

[0164] A target data feature is generated according to the discriminant feature and the confidence value.

[0165] In the present invention, for a data feature extraction device, first, the present invention obtains an initial image to be detected and extracts an initial feature sequence of the initial image to be detected. The multi-layer convolution layer can capture the rich structure and texture information in the image at multiple scales and levels. The average pooling effectively reduces the feature dimension, enhances the translation invariance of the feature, and improves the computational efficiency. The activation function introduces nonlinear transformation, enhances the feature expression capability, and highlights the key areas. The context dependency of the initial feature sequence is captured and encoded to obtain a number of coding sequence features. Then, the bidirectional capture model is used to perform forward and reverse recursive calculations and aggregation on the initial feature sequence, fully exploring the context dependency between adjacent and distant elements in the sequence, improving the semantic expression capability of the feature, determining the importance score of each coding sequence feature, and dynamically screening the importance score according to a preset adaptive threshold to obtain the key sequence features corresponding to the screened importance score. The coding sequence features are assigned a weight vector and the dot product score is calculated to quantify the importance of each feature. After normalization mapping, the scores are ensured to be comparable. A screening strategy based on a preset adaptive threshold and a dynamic weighted smoothing mechanism is adopted to effectively balance the influence of historical thresholds and current scores, thereby achieving dynamic identification and filtering of key sequence features. The key sequence features are pyramid fused to obtain fused sequence features. The convolution features are spatially aligned using bilinear interpolation to ensure the consistency of features of different scales in the spatial dimension. The multi-scale aligned features are fused through a splicing operation to enrich the diversity and integrity of feature expression. Finally, the initial image to be detected is feature recognized based on the fused sequence features to obtain target data features. The target data features are generated by combining the discriminant features with the confidence level, which not only reflects the key information of the image but also includes the confidence level of the results, thereby effectively improving the accuracy and efficiency of data feature extraction. The specific definition of a data feature extraction device can be found in the definition of a data feature extraction method above, which will not be repeated here. The various modules in the above-mentioned data feature extraction device can be implemented in whole or in part by software, hardware, and a combination thereof. The above modules may be embedded in or independent of the processor in the computer device in the form of hardware, or may be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0166] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 5 As shown. The computer device includes a processor, memory, network interface and database connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external client via a network connection. When the computer program is executed by the processor, it implements the functions or steps on the server side of a data feature extraction method.

[0167] In one embodiment, a computer device is provided. The computer device may be a client, and its internal structure diagram may be as follows: Figure 6 As shown. The computer device includes a processor, memory, network interface, display screen, and input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements the functions or steps of the client side of a data feature extraction method.

[0168] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the following steps are performed:

[0169] Acquire an initial image to be detected, and extract an initial feature sequence of the initial image to be detected;

[0170] Capturing and encoding context dependencies of the initial feature sequence to obtain a plurality of coded sequence features;

[0171] Determining the importance score of each of the coding sequence features, and dynamically screening the importance scores according to a preset adaptive threshold to obtain key sequence features corresponding to the screened importance scores;

[0172] Performing pyramid fusion on the key sequence features to obtain fused sequence features;

[0173] Feature recognition is performed on the initial image to be detected according to the fusion sequence features to obtain target data features.

[0174] In the several embodiments provided by the present invention, it should be understood that the disclosed devices and apparatuses can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the module division is merely a logical function division, and actual implementation may employ other division methods.

[0175] In addition, the functional modules in various embodiments of the present invention may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or hardware plus software functional modules.

[0176] Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims rather than the foregoing description, and all changes that come within the meaning and range of equivalents of the claims are intended to be embraced therein. Any reference to a figure in a claim should not be construed as limiting the claim to which it relates.

[0177] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0178] In some implementations of this embodiment, a computer-readable storage medium is provided, on which a computer program is stored, characterized in that when the computer program is executed by a processor, the steps of the method described in the above embodiment are implemented.

[0179] The readable storage medium of the present invention stores a computer program, which, when executed by a processor of an electronic device, can implement:

[0180] Acquire an initial image to be detected, and extract an initial feature sequence of the initial image to be detected;

[0181] Capturing and encoding context dependencies of the initial feature sequence to obtain a plurality of coded sequence features;

[0182] Determining the importance score of each of the coding sequence features, and dynamically screening the importance scores according to a preset adaptive threshold to obtain key sequence features corresponding to the screened importance scores;

[0183] Performing pyramid fusion on the key sequence features to obtain fused sequence features;

[0184] Feature recognition is performed on the initial image to be detected according to the fusion sequence features to obtain target data features.

[0185] It should be noted that the above functions or steps that can be implemented by the computer-readable storage medium or computer device can be found in the relevant descriptions of the server side and the client side in the aforementioned method embodiment. To avoid repetition, they will not be described one by one here.

[0186] The computer-readable storage medium may also store at least one computer-executable program / instruction, such as a computer-readable instruction. Computer-readable storage media include, but are not limited to, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory. Computer-readable storage media may include, for example, read-only memory (ROM), a hard disk, a flash memory, etc. For example, a non-transitory computer-readable storage medium may be connected to a computing device such as a computer, and then, when the computing device executes the computer-readable instructions stored on the computer-readable storage medium, the various methods described above may be performed.

[0187] In addition, the computer device may also include (but is not limited to) a data bus, an input / output (I / O) bus, a display, and input / output devices (eg, keyboard, mouse, speaker, etc.).

[0188] The processor can communicate with external devices via an I / O bus via a wired or wireless network.

[0189] In one embodiment, the at least one computer executable instruction may also be compiled into or constitute a software product / computer program product, wherein one or more computer executable instructions are executed by a processor to perform the various functions and / or method steps in the embodiments described in the present technology.

[0190] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0191] Those skilled in the art will clearly understand that for the sake of convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0192] In the embodiments provided in the present disclosure, it should be understood that the disclosed devices and methods can also be implemented in other ways. The device embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architectures, functions and operations of the devices, methods and computer program products according to multiple embodiments of the present disclosure. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a portion of code, and the above-mentioned module, program segment or a portion of code contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or can be implemented using a combination of dedicated hardware and computer instructions.

[0193] It should be noted that, in this disclosure, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element limited by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or apparatus comprising the element.

[0194] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention, and should all be included in the scope of protection of the present invention.

[0195] It should be noted that if software tools or components other than those of our company appear in the embodiments of this application, they are only used for illustration and do not represent actual use.

Claims

1. A data feature extraction method, characterized in that: The method comprises: Acquire an initial image to be detected, and extract an initial feature sequence of the initial image to be detected; Capturing and encoding context dependencies of the initial feature sequence to obtain a plurality of coded sequence features; Determining the importance score of each of the coding sequence features, and dynamically screening the importance scores according to a preset adaptive threshold to obtain key sequence features corresponding to the screened importance scores; Performing pyramid fusion on the key sequence features to obtain fused sequence features; Feature recognition is performed on the initial image to be detected according to the fusion sequence features to obtain target data features.

2. The data feature extraction method according to claim 1, wherein: The extracting of the initial feature sequence of the initial image to be detected includes: Performing denoising processing on the initial image to be detected to obtain a denoised image; Convolving the denoised image to obtain a convolved image; Performing average pooling on the convolution image to obtain a pooled image; Activating the pooled image using a preset activation function to obtain spatial local features; A spatial dimension is obtained, and the spatial local features are mapped into an initial feature sequence according to the spatial dimension.

3. The data feature extraction method according to claim 1, wherein: The context dependency is captured and encoded on the initial feature sequence to obtain a plurality of encoding sequence features, including: Recursively calculating and summarizing the initial feature sequence in a preset first direction and a preset second direction using a preset bidirectional capture model to obtain a hidden state sequence group; splicing the sequences in the hidden state sequence group to obtain context-enhanced features; Non-linear activation is performed on the context enhancement feature to obtain a coding sequence feature.

4. The data feature extraction method according to claim 1, wherein: Determining the importance score of each coding sequence feature comprises: Obtaining a weight vector for each of the coding sequence features, performing a dot product between each of the coding sequence features and the corresponding weight vector to obtain a dot product score; The point product scores are normalized and mapped to obtain importance scores.

5. The data feature extraction method according to claim 1, wherein: The dynamically screening the importance scores according to the preset adaptive threshold to obtain key sequence features corresponding to the screened importance scores includes: Obtaining an initial threshold value of a preset time step and an importance score of a next time step of the preset time step; Obtaining a smoothing coefficient, and generating a weighted weight according to the smoothing coefficient; Performing a weighted average of the initial threshold value and the importance score of the next time step using the weighted weight to obtain an updated threshold value; Determining whether the importance score of the next time step is greater than the threshold update value; If the importance score of the next time step is less than or equal to the threshold update value, deleting the coding sequence feature corresponding to the importance score of the next time step; If the importance score of the next time step is greater than the threshold update value, filtering out the coding sequence feature corresponding to the importance score of the next time step, and confirming the activation attention of the filtered coding sequence feature; Key sequence features are generated based on the activated attention and the filtered coding sequence features.

6. The data feature extraction method according to claim 1, wherein: The pyramid fusion of the key sequence features to obtain fused sequence features includes: Performing multi-layer convolution on the key sequence features to obtain several key convolution features; Performing bilinear interpolation on the key convolution features to obtain aligned convolution features; Splicing the aligned convolution features to obtain spliced ​​convolution features; A scaling parameter and a translation parameter are obtained, and the concatenated convolution features are normalized using the scaling parameter and the translation parameter to obtain a fused sequence feature.

7. The data feature extraction method according to claim 1, wherein: The performing feature recognition on the initial image to be detected according to the fusion sequence features to obtain target data features includes: Acquire a target analysis task, map the fused sequence features to a target space according to the target analysis task, and extract discriminant features of the initial image to be detected in the target space; Determining a confidence value of the target analysis task based on the discriminant feature; A target data feature is generated according to the discriminant feature and the confidence value.

8. A data feature extraction device, characterized in that: The device comprises: A feature sequence extraction module is used to obtain an initial image to be detected and extract an initial feature sequence of the initial image to be detected; A dependency capturing module, configured to capture and encode context dependencies of the initial feature sequence to obtain a plurality of encoding sequence features; A key feature screening module is used to determine the importance score of each coding sequence feature and dynamically screen the importance score according to a preset adaptive threshold to obtain the key sequence feature corresponding to the screened importance score; A key feature fusion module is used to perform pyramid fusion on the key sequence features to obtain fused sequence features; The feature recognition module is used to perform feature recognition on the initial image to be detected based on the fusion sequence features to obtain target data features.

9. An electronic device, characterized in that: The electronic device comprises: at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute a data feature extraction method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, a data feature extraction method according to any one of claims 1 to 7 is implemented.

Images