A multi-stage feature processing method and related device

Abstract

This invention provides a multi-stage feature processing method and related equipment, relating to the field of image processing technology. By introducing a post-feature encoding module, unified encoding and adaptive channel compression of output features are achieved, reducing feature redundancy and improving the standardization and stability of feature representation. By introducing a global feature reweighting reconstruction module, the global correlation between different spatial locations is explicitly modeled, enabling output features to integrate global contextual information, significantly enhancing the overall consistency and structural expressive power of features. By setting a reconstructed feature decoding output module, the nonlinear modeling and structured decoding capabilities of the output stage are improved, making the prediction results more continuous, stable, and refined in complex background scenes.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a multi-stage feature processing method and related equipment. Background Technology

[0002] With the rapid development of deep learning and computer vision technologies, image analysis and understanding methods based on deep neural networks have been widely applied in various fields such as image segmentation, object detection, semantic understanding, and scene parsing. These methods typically construct specific deep computing models to extract and learn features from the input image layer by layer, and then model and infer target information in the feature space. The structural design of the computing model, the feature representation method, and the output inference mechanism directly affect the model's expressive power, inference stability, and practical application effectiveness in complex scenes.

[0003] In existing technologies, encoder-decoder deep network structures are widely used in various image understanding tasks due to their good performance in multi-scale feature fusion and spatial resolution restoration. However, with the continuous increase in the complexity of application scenarios and the requirements for accuracy, existing computational models have gradually revealed certain limitations in areas such as unified feature modeling, global dependency representation, and feature reconstruction capabilities in the output stage.

[0004] Existing deep learning-based image processing computational models typically include a front-end feature extraction module and a back-end output inference module. The front-end feature extraction module generally consists of multiple convolutional operations, used to extract low-level to high-level feature representations of the image layer by layer. The back-end output inference module usually generates the final prediction result based on the features generated by the front-end through simple mapping or regression. In computational models that partially adopt an encoder-decoder structure, the encoding stage extracts high-level semantic information through multiple downsampling and feature transformations, while the decoding stage restores spatial resolution through upsampling and feature fusion, generating output features consistent with the input size. This type of model structure has achieved multi-scale information fusion and spatial structure restoration to a certain extent, becoming the mainstream technical approach in current image processing tasks. However, in the above computational flow, the features output by the decoding stage are often directly used for result prediction, and their internal feature representation and inference mechanism are relatively simple, lacking further unified encoding, global relationship modeling, and fine reconstruction processing for the output features.

[0005] From the perspective of feature representation, existing computational models typically output features with high spatial resolution and large channel dimensions during the decoding stage. However, the semantic distribution between different channels is relatively scattered, lacking a unified constraint and standardized feature representation mechanism. When this feature form is directly used for result inference, it is easily affected by local noise or feature redundancy, which is detrimental to the stability and consistency of the model output. From the perspective of global modeling capability, existing computational models mainly rely on local convolutional operations for feature modeling. Their receptive field expansion depends on network depth or multi-scale structure design, making it difficult to display global correlations between arbitrary spatial locations of the client. In complex scenarios, the model's ability to express long-distance dependencies, regional consistency, and overall structural constraints is limited, which can easily lead to global inconsistencies in the output results. In terms of semantic consistency, existing models are insufficient. From the perspective of feature reconstruction and inference mechanisms, they typically use a single mapping method to predict features in the output stage, lacking a reconstruction and reweighting process for output features. This fails to fully explore the correlation and complementary relationships between features. When faced with application scenarios with complex boundaries, high requirements for structural continuity, or significant contextual correlations, this direct output method is prone to problems such as unsmooth prediction results or local inconsistencies. From the perspective of model structure scalability, existing computational models often tightly couple feature extraction with the output inference process. The output stage lacks an independent structured processing module, which limits the transfer and reuse capabilities of the model across different interfaces or different front-end network structures, hindering the flexible expansion of the model structure and its engineering applications.

[0006] In summary, existing deep learning-based image processing computing models still have shortcomings in terms of unified encoding of output features, global correlation modeling capabilities, and feature reconstruction and inference mechanisms, making it difficult to simultaneously ensure the accuracy, stability, and structural consistency of output results in complex scenarios. Summary of the Invention

[0007] This invention provides a multi-stage feature processing method and related equipment, the purpose of which is to improve the accuracy and stability of the output results.

[0008] To achieve the above objectives, the present invention provides a multi-stage feature processing method, comprising:

[0009] Step 1: Obtain the training image dataset;

[0010] Step 2: Train the constructed multi-stage feature reconstruction model using the training image dataset to obtain the trained multi-stage feature reconstruction model.

[0011] Step 3: Input the image data to be processed into the trained multi-stage feature reconstruction model for processing to obtain the multi-stage feature processing results;

[0012] The multi-stage feature reconstruction model includes a feature generation module for multi-level feature extraction and modeling of the input data, a post-feature encoding module for encoding and compressing the initial features output by the feature generation module, a global feature reweighting reconstruction module for weighted reconstruction of the encoded features output by the post-feature encoding module, and a reconstructed feature decoding output module for multi-stage decoding and mapping of the reconstructed features output by the global feature reweighting reconstruction module.

[0013] Furthermore, before processing the image data to be processed by inputting it into the trained multi-stage feature reconstruction model, the following steps are also included:

[0014] The image data to be processed is preprocessed to obtain preprocessed image data.

[0015] Furthermore, the post-feature encoding module includes multi-branch convolutional units, fusion units, and channel compression units;

[0016] The input of the multi-branch convolutional unit is connected to the output of the feature generation module, and the output of the multi-branch convolutional unit is connected to the input of the fusion unit.

[0017] The output of the fusion unit is connected to the input of the channel compression unit, and the output of the channel compression unit is connected to the input of the global feature weight reconstruction module.

[0018] Furthermore, the channel compression unit includes a global average pooling layer, a multilayer perceptron, a convolutional layer, and a group normalization layer;

[0019] The input of the global average pooling layer is connected to the output of the fusion unit;

[0020] The first output of the global average pooling layer is connected to the input of the multilayer perceptron;

[0021] The second output of the global average pooling layer and the output of the multilayer perceptron are both connected to the input of the convolutional layer.

[0022] The output of the convolutional layer is connected to the input of the group normalization layer, and the output of the group normalization layer is connected to the input of the global feature weight reconstruction module.

[0023] Furthermore, the global feature reweighting and reconstruction module includes a feature mapping unit, a global association weight calculation unit, a feature reconstruction unit, and a feature fusion unit;

[0024] The input of the feature mapping unit and the output of the group normalization layer;

[0025] The first output of the feature mapping unit is connected to the input of the global correlation weight calculation unit;

[0026] The second output of the feature mapping unit and the output of the global correlation weight calculation unit are both connected to the input of the feature reconstruction unit;

[0027] The output of the feature reconstruction unit and the output of the group normalization layer are both connected to the input of the feature fusion unit.

[0028] Furthermore, the global correlation weight calculation unit is used to calculate the correlation weight between any spatial locations, and the calculation expression is:

[0029] ;

[0030] in, Indicates spatial location The correlation weight between them Indicates the attention head index. Indicates the number of intermediate feature channels. Indicates the first A position in the attention. The query feature vector, Indicates the first A position in the attention. The key feature vector.

[0031] The present invention also provides a multi-stage feature processing apparatus, comprising:

[0032] The acquisition module is used to acquire the image dataset for training.

[0033] The training module is used to train the constructed multi-stage feature reconstruction model using the training image dataset to obtain the trained multi-stage feature reconstruction model.

[0034] The processing module is used to input the image data to be processed into the trained multi-stage feature reconstruction model for processing, and obtain the multi-stage feature processing results.

[0035] The multi-stage feature reconstruction model includes a feature generation module for multi-level feature extraction and modeling of the input data, a post-feature encoding module for encoding and compressing the initial features output by the feature generation module, a global feature reweighting reconstruction module for weighted reconstruction of the encoded features output by the post-feature encoding module, and a reconstructed feature decoding output module for multi-stage decoding and mapping of the reconstructed features output by the global feature reweighting reconstruction module.

[0036] The present invention also provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a multi-stage feature processing method.

[0037] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a multi-stage feature processing method.

[0038] The above-described solution of the present invention has the following beneficial effects:

[0039] This invention trains a multi-stage feature reconstruction model using a training image dataset to obtain the trained multi-stage feature reconstruction model. The image data to be processed is then input into the trained multi-stage feature reconstruction model for processing, yielding multi-stage feature processing results. The multi-stage feature reconstruction model includes, in sequence, a feature generation module for multi-level feature extraction and modeling of the input data, a post-feature encoding module for encoding and compressing the initial features output by the feature generation module, a global feature reweighting reconstruction module for weighted reconstruction of the encoded features output by the post-feature encoding module, and a multi-stage feature reweighting module for multi-level reconstructing of the reconstructed features output by the global feature reweighting reconstruction module. The present invention provides a segment decoding and mapping reconstruction feature decoding output module. Compared with existing technologies, this invention introduces a post-feature encoding module to achieve unified encoding and adaptive channel compression of output features, reducing feature redundancy and improving the standardization and stability of feature representation. By introducing a global feature reweighting reconstruction module, the global correlation between different spatial locations is explicitly modeled, enabling output features to integrate global context information, significantly enhancing the overall consistency and structural expressive power of features. By setting up a reconstruction feature decoding output module, the nonlinear modeling and structured decoding capabilities of the output stage are improved, making the prediction results more continuous, stable, and refined in complex background scenarios.

[0040] Other beneficial effects of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0041] Figure 1 This is a flowchart illustrating an embodiment of the present invention;

[0042] Figure 2 This is a schematic diagram of the structure of the multi-stage feature reconstruction model in an embodiment of the present invention;

[0043] Figure 3 This is a schematic diagram of the structure of the multi-stage feature processing device in an embodiment of the present invention;

[0044] Figure 4 This is a schematic diagram of the structure of the terminal device in an embodiment of the present invention. Detailed Implementation

[0045] To make the technical problems, solutions, and advantages of this invention clearer, a detailed description will be provided below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0046] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a locking connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0047] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0048] This invention addresses existing problems by providing a multi-stage feature processing method and related equipment.

[0049] like Figure 1 , Figure 2 As shown, an embodiment of the present invention provides a multi-stage feature processing method, including:

[0050] Step 1: Obtain the training image dataset;

[0051] Step 2: Train the constructed multi-stage feature reconstruction model using the training image dataset to obtain the trained multi-stage feature reconstruction model.

[0052] Step 3: Input the image data to be processed into the trained multi-stage feature reconstruction model for processing to obtain the multi-stage feature processing results;

[0053] The multi-stage feature reconstruction model includes a feature generation module for multi-level feature extraction and modeling of the input data, a post-feature encoding module for encoding and compressing the initial features output by the feature generation module, a global feature reweighting reconstruction module for weighted reconstruction of the encoded features output by the post-feature encoding module, and a reconstructed feature decoding output module for multi-stage decoding and mapping of the reconstructed features output by the global feature reweighting reconstruction module.

[0054] Preferably, before inputting the image data to be processed into the trained multi-stage feature reconstruction model, the following steps are also included:

[0055] The image data to be processed is preprocessed to obtain preprocessed image data.

[0056] In this embodiment of the invention, the preprocessing of the image data to be processed includes, but is not limited to, size unification, normalization, noise reduction, or data format conversion.

[0057] In the embodiments of the present invention, the feature generation module may be a convolutional neural network, a visual transformation module, or other models capable of outputting spatial feature representations, and its specific structure does not constitute a limitation on the embodiments of the present invention.

[0058] To further improve the stability and robustness of subsequent feature encoding and reconstruction processes, the initial features output by the feature generation module are subjected to the following operations in this embodiment of the invention:

[0059] Channel standardization: Normalize or scale the initial features along the channel dimension to keep the numerical distribution of different channels within a controllable range;

[0060] Noise reduction: Reduce potential noise interference in the initial verification by smoothing, filtering or suppressing abnormal responses;

[0061] Size adjustment processing: Based on the calculation requirements of subsequent modules, the features are downsampled or upsampled to match a uniform spatial size.

[0062] Since the initial features output by the feature generation module usually have a large number of channels and a scattered semantic distribution, if they are directly used for global modeling or output mapping, they are prone to introducing redundant calculations and amplifying noise response. In this embodiment of the invention, a post-feature encoding module is introduced, which includes a multi-branch convolutional unit, a fusion unit, and a channel compression unit.

[0063] The input of the multi-branch convolutional unit is connected to the output of the feature generation module, and the output of the multi-branch convolutional unit is connected to the input of the fusion unit.

[0064] The output of the fusion unit is connected to the input of the channel compression unit, and the output of the channel compression unit is connected to the input of the global feature weight reconstruction module.

[0065] In order to achieve effective compression of channel dimensions and unified semantic representation of output features while maintaining the original spatial resolution, this embodiment of the invention utilizes multi-branch convolutional units in the post-feature encoding module to process the processed features in parallel to capture local information at different scales, and then uses a fusion unit to fuse the multi-branch features. Subsequently, the normalization processing layer and nonlinear activation function in the fusion module are used to stabilize the fused features. Finally, the high-dimensional features are mapped into low-dimensional, compact, and semantically more concentrated encoded features through the channel compression unit.

[0066] Specifically, a multi-branch convolutional unit consists of three sets of parallel convolutional layers of different sizes, a group normalization layer, and a non-linear activation function, calculated as follows:

[0067] ;

[0068] ;

[0069] ;

[0070] in, Indicates the processed features. , Indicates the kernel size as Two-dimensional convolution operation, This indicates a group normalization operation, used to enhance training stability. This represents a non-linear activation function; in this embodiment of the invention, it is the ReLU activation function. Indicates local details, Represents scale context features, This represents channel statistics and semantic compression features.

[0071] Specifically, the fusion unit consists of splicing layers, and the calculation expression is as follows:

[0072] ;

[0073] in, Indicates fusion characteristics, This indicates feature splicing along the channel dimension.

[0074] Specifically, the channel compression unit includes a global average pooling layer, a multilayer perceptron, a convolutional layer, and a group normalization layer;

[0075] The input of the global average pooling layer is connected to the output of the fusion unit;

[0076] The first output of the global average pooling layer is connected to the input of the multilayer perceptron;

[0077] The second output of the global average pooling layer and the output of the multilayer perceptron are both connected to the input of the convolutional layer.

[0078] The output of the convolutional layer is connected to the input of the group normalization layer, and the output of the group normalization layer is connected to the input of the global feature weight reconstruction module.

[0079] In this embodiment of the invention, the calculation expression for the channel compression unit is:

[0080] ;

[0081] ;

[0082] in, Represents coding features, , This indicates a global average pooling operation. This represents a multilayer perceptron used to learn the channel weight mapping relationship. This represents the sigmoid activation function. This represents the channel adaptive weight vector. This indicates channel-by-channel multiplication. This indicates channel compression and unified encoding.

[0083] Although the encoded features processed by the post-feature encoding module have good uniformity, they still mainly reflect local spatial information and are difficult to express the global dependency relationship between different spatial locations. Therefore, this embodiment of the invention introduces a global feature reweighting reconstruction module, which includes a feature mapping unit, a global association weight calculation unit, a feature reconstruction unit, and a feature fusion unit.

[0084] The input of the feature mapping unit and the output of the group normalization layer;

[0085] The first output of the feature mapping unit is connected to the input of the global correlation weight calculation unit;

[0086] The second output of the feature mapping unit and the output of the global correlation weight calculation unit are both connected to the input of the feature reconstruction unit;

[0087] The output of the feature reconstruction unit and the output of the group normalization layer are both connected to the input of the feature fusion unit.

[0088] In this embodiment of the invention, the global feature weighting reconstruction module maps the input encoded features through the feature mapping unit, and then calculates the similarity or correlation of the encoded features at different spatial locations through the global association weight calculation unit to obtain the correlation weight. Then, in the feature reconstruction unit, the correlated weight is used to perform weighted fusion of the encoded features so that the features at each spatial location can incorporate contextual information from the global scope. Finally, in the feature fusion unit, the weighted fused features are superimposed with the encoded features using a residual fusion method to preserve the original local information and enhance the stability of the training and inference process.

[0089] Specifically, the expression for mapping the input encoded features through the feature mapping unit is as follows:

[0090] ;

[0091] ;

[0092] ;

[0093] in, Indicates the first The query feature vector of each attention head, Indicates the first The key feature vector of each attention head, Indicates the first The value feature vector of each attention head, , , denoted as the learnable convolutional kernel weight matrices used to generate query, key, and value feature vectors, respectively.

[0094] Specifically, the global correlation weight calculation unit is used to calculate the correlation weight between any spatial locations, and the calculation expression is:

[0095] ;

[0096] in, Indicates spatial location The correlation weight between them Indicates the attention head index. Indicates the number of intermediate feature channels. Indicates the first A position in the attention. The query feature vector, Indicates the first A position in the attention. The key feature vector.

[0097] Specifically, the calculation expression for weighted fusion of encoded features using relevance weights in the feature reconstruction unit is as follows:

[0098] ;

[0099] in, Indicates the first Weighted fusion features of individual attention heads Indicates the first A position in the attention. The eigenvectors of the values.

[0100] Specifically, the calculation expression for superimposing reconstructed features and encoded features using residual fusion in the feature fusion unit is as follows:

[0101] ;

[0102] in, Represents the reconstructed features, This represents the learnable residual fusion coefficient.

[0103] Specifically, the reconstructed feature decoding output module, used for multi-stage decoding and mapping of the reconstructed features output by the global feature reweighting reconstruction module, includes:

[0104] Introducing a phased reconstruction feature decoding operation, the calculation expression is:

[0105] ;

[0106] ;

[0107] ;

[0108] in, This indicates that a channel projection operation is performed on the reconstructed features;

[0109] The decoded features are mapped, and the expression is:

[0110] ;

[0111] in, This represents the output of the multi-stage feature processing results. This represents the learnable output fusion weights. Used to map the intermediate decoded features to the number of task output channels.

[0112] This invention utilizes a training image dataset to train a constructed multi-stage feature reconstruction model, resulting in a trained multi-stage feature reconstruction model. The image data to be processed is then input into the trained multi-stage feature reconstruction model for processing, yielding multi-stage feature processing results. The multi-stage feature reconstruction model includes, in sequence, a feature generation module for multi-level feature extraction and modeling of the input data, a post-feature encoding module for encoding and compressing the initial features output by the feature generation module, a global feature reweighting reconstruction module for weighted reconstruction of the encoded features output by the post-feature encoding module, and a multi-stage feature reweighting module for multi-level reconstructing of the reconstructed features output by the global feature reweighting reconstruction module. The present invention provides a segment decoding and mapping reconstruction feature decoding output module. Compared with the prior art, this embodiment of the invention introduces a post-feature encoding module to achieve unified encoding and channel adaptive compression of output features, reducing feature redundancy and improving the standardization and stability of feature representation. By introducing a global feature reweighting reconstruction module, the global correlation between different spatial locations is explicitly modeled, enabling the output features to integrate global context information, significantly enhancing the overall consistency and structural expressive power of the features. By setting up a reconstruction feature decoding output module, the nonlinear modeling and structured decoding capabilities of the output stage are improved, making the prediction results more continuous, stable and refined in complex background scenarios.

[0113] Corresponding to the multi-stage feature processing method described in the above embodiments, such as Figure 3 As shown, this embodiment of the invention also provides a multi-stage feature processing device 100, which includes:

[0114] Module 101 is used to acquire the training image dataset;

[0115] Training module 102 is used to train the constructed multi-stage feature reconstruction model using the training image dataset to obtain the trained multi-stage feature reconstruction model.

[0116] Processing module 103 is used to input the image data to be processed into the trained multi-stage feature reconstruction model for processing, and obtain multi-stage feature processing results;

[0117] The multi-stage feature reconstruction model includes a feature generation module for multi-level feature extraction and modeling of the input data, a post-feature encoding module for encoding and compressing the initial features output by the feature generation module, a global feature reweighting reconstruction module for weighted reconstruction of the encoded features output by the post-feature encoding module, and a reconstructed feature decoding output module for multi-stage decoding and mapping of the reconstructed features output by the global feature reweighting reconstruction module.

[0118] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0119] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to 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. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0120] This invention also provides a terminal device, such as... Figure 4 As shown, the terminal device D10 of this embodiment includes: at least one processor D100 ( Figure 4 The diagram shows only one processor, a memory D101, and a computer program D102 stored in the memory D101 and executable on the at least one processor D100. When the processor D100 executes the computer program D102, it implements the multi-stage feature processing method described above.

[0121] The terminal device D10 can be a desktop computer, laptop, handheld computer, server, server cluster, or cloud server, etc. This terminal device may include, but is not limited to, a processor D100 and a memory D101. Those skilled in the art will understand that... Figure 4 This is merely an example of terminal device D10 and does not constitute a limitation on terminal device D10. It may include more or fewer components than shown in the figure, or combine certain components, or different components, such as input / output devices, network access devices, etc.

[0122] The processor D100 can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0123] In some embodiments, the memory D101 may be an internal storage unit of the terminal device D10, such as a hard disk or memory of the terminal device D10. In other embodiments, the memory D101 may be an external storage device of the terminal device D10, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the terminal device D10. Furthermore, the memory D101 may include both internal and external storage units of the terminal device D10. The memory D101 is used to store the operating system, applications, bootloader, data, and other programs, such as the program code of the computer program. The memory D101 can also be used to temporarily store data that has been output or will be output.

[0124] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0125] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to 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. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0126] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a multi-stage feature processing method.

[0127] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying the computer program code to a building device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0128] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A multi-stage feature processing method, characterized by, The method comprises the following steps: Step 1, obtaining a training image data set; Step 2, training a multi-stage feature reconstruction model constructed by using the training image data set to obtain a trained multi-stage feature reconstruction model; Step 3, inputting image data to be processed into the trained multi-stage feature reconstruction model for processing to obtain a multi-stage feature processing result; The multi-stage feature reconstruction model comprises a feature generation module connected in sequence for multi-layer feature extraction and modeling of input data, a post-feature encoding module for encoding and compressing initial features output by the feature generation module, a global feature reweighting reconstruction module for weighting and reconstructing encoded features output by the post-feature encoding module, and a reconstructed feature decoding output module for multi-stage decoding and mapping of reconstructed features output by the global feature reweighting reconstruction module; The post-feature encoding module comprises a multi-branch convolution unit for parallel processing of processed features, a fusion unit for fusing multi-branch features, and a channel compression unit for mapping high-dimensional features into encoded features; The input end of the multi-branch convolution unit is connected with the output end of the feature generation module, and the output end of the multi-branch convolution unit is connected with the input end of the fusion unit; The output end of the fusion unit is connected with the input end of the channel compression unit, and the output end of the channel compression unit is connected with the input end of the global feature reweighting reconstruction module; The global feature reweighting reconstruction module comprises a feature mapping unit for mapping input encoded features, a global correlation weight calculation unit for calculating the similarity or correlation of encoded features at different spatial positions, a feature reconstruction unit for weighting and fusing encoded features by using correlation weights, and a feature fusion unit for superimposing weighted and fused features and encoded features; The input end of the feature mapping unit is connected with the output end of the group normalization layer in the channel compression unit; The first output end of the feature mapping unit is connected with the input end of the global correlation weight calculation unit; The second output end of the feature mapping unit and the output end of the global correlation weight calculation unit are both connected with the input end of the feature reconstruction unit; The output end of the feature reconstruction unit and the output end of the group normalization layer in the channel compression unit are both connected with the input end of the feature fusion unit.

2. The multi-stage feature processing method of claim 1, wherein, Before inputting the image data to be processed into the trained multi-stage feature reconstruction model for processing, the method further comprises the following steps: Pretreating the image data to be processed to obtain pretreated image data.

3. The multi-stage feature processing method of claim 1, wherein, The channel compression unit comprises a global average pooling layer, a multilayer perceptron, a convolution layer, and a group normalization layer; The input end of the global average pooling layer is connected with the output end of the fusion unit; The first output end of the global average pooling layer is connected with the input end of the multilayer perceptron; The second output end of the global average pooling layer and the output end of the multilayer perceptron are both connected with the input end of the convolution layer; The output of the convolutional layer is connected to the input of the group normalization layer, and the output of the group normalization layer is connected to the input of the global feature weight reconstruction module.

4. The multi-stage feature processing method of claim 1, wherein, The global correlation weight calculation unit is used to calculate the correlation weight between arbitrary spatial locations, and the calculation expression is: ； wherein, denotes a spatial position, a correlation weight between, denotes an attention head index, denotes an intermediate feature channel number, denotes a query feature vector at a position in the th attention head, denotes a key feature vector at a position in the th attention head.

5. A multi-stage feature processing apparatus characterized by comprising: include: The acquisition module is used to acquire the image dataset for training. The training module is used to train the constructed multi-stage feature reconstruction model using the training image dataset to obtain the trained multi-stage feature reconstruction model. The processing module is used to input the image data to be processed into the trained multi-stage feature reconstruction model for processing, and obtain the multi-stage feature processing results. The multi-stage feature reconstruction model includes a feature generation module for multi-level feature extraction and modeling of input data, a post-feature encoding module for encoding and compressing the initial features output by the feature generation module, a global feature reweighting reconstruction module for weighted reconstruction of the encoded features output by the post-feature encoding module, and a reconstruction feature decoding output module for multi-stage decoding and mapping of the reconstructed features output by the global feature reweighting reconstruction module. The post-feature encoding module includes a multi-branch convolutional unit for parallel processing of the processed features, a fusion unit for fusing the multi-branch features, and a channel compression unit for mapping high-dimensional features to encoded features. The input of the multi-branch convolutional unit is connected to the output of the feature generation module, and the output of the multi-branch convolutional unit is connected to the input of the fusion unit. The output of the fusion unit is connected to the input of the channel compression unit, and the output of the channel compression unit is connected to the input of the global feature weight reconstruction module. The global feature weighting reconstruction module includes a feature mapping unit for mapping the input encoded features, a global association weight calculation unit for calculating the similarity or correlation of encoded features at different spatial locations, a feature reconstruction unit for weighted fusion of encoded features using correlation weights, and a feature fusion unit for superimposing the weighted fused features with the encoded features. The input of the feature mapping unit is connected to the output of the group normalization layer in the channel compression unit; The first output of the feature mapping unit is connected to the input of the global correlation weight calculation unit; The second output of the feature mapping unit and the output of the global correlation weight calculation unit are both connected to the input of the feature reconstruction unit. The output of the feature reconstruction unit and the output of the group normalization layer in the channel compression unit are both connected to the input of the feature fusion unit.

6. A terminal device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the multi-stage feature processing method as described in any one of claims 1 to 4.

7. A computer-readable storage medium storing a computer program, wherein the computer program comprises the following steps of: receiving a request for a resource from a client; determining whether the client is authorized to access the resource; and if the client is authorized to access the resource, providing the resource to the client. When the computer program is executed by the processor, it implements the multi-stage feature processing method as described in any one of claims 1 to 4.

Images