Feature Enhancement Method, Apparatus, Data Processing Method, and Electronic Device

Abstract

The present application relates to a feature enhancement method, an apparatus, a data processing method, and an electronic device. Among them, in the feature enhancement method, a first feature map and a second feature map having the same size as the first feature map obtained after preprocessing the first feature map are acquired. The first feature map is a feature map obtained by performing feature extraction on a target image. The first feature map is processed using the DCA algorithm to obtain a first transformed feature map, and the second feature map is processed using the DCA algorithm to obtain a second transformed feature map. The first transformed feature map is processed using the SE-Net algorithm, and the processing result is fused with the second transformed feature map, and an enhanced feature map is obtained based on the fusion result. It utilizes a network structure based on SE-Net and DCA, and fuses image features according to the correlation information between pixel points and the correlation between structures in the dataset, thereby improving the accuracy of image recognition based on the fused features.

Description

Technical Field

[0001] The present application relates to the field of data, and particularly to a feature enhancement method, an apparatus, a data processing method, and an electronic device. Background Art

[0002] With the development of technology and the improvement of the economic level, currently, image recognition technology has been widely applied in fields such as traffic management, household equipment, and security, penetrating into all aspects of people's production and life, and gradually building an intelligent society.

[0003] Currently, convolutional neural networks have been deeply studied and developed, and have played an important role in the field of image recognition technology. In image recognition technology, in order to improve the performance of convolutional neural networks, image features can be enhanced through feature fusion. However, current feature fusion methods often simply add the pixels at corresponding positions directly on the feature map, lacking the utilization of the correlation information between pixel points and the correlation between structures in the dataset, resulting in low accuracy of image recognition.

[0004] Regarding the problem in the related technology that there is a lack of utilization of the correlation information between pixel points and the correlation between structures in the dataset during the feature fusion process, resulting in low accuracy of image recognition, no effective solution has been proposed yet. Summary of the Invention

[0005] In this embodiment, a feature enhancement method, an apparatus, a data processing method, and an electronic device are provided to solve the problem in the related technology that there is a lack of utilization of the correlation information between pixel points and the correlation between structures in the dataset during the feature fusion process, resulting in low accuracy of image recognition.

[0006] In a first aspect, in this embodiment, a feature enhancement method for image recognition is provided, including:

[0007] Obtain a first feature map and a second feature map of the same size as the first feature map obtained after preprocessing the first feature map; wherein, the first feature map is a feature map obtained by performing feature extraction on a target image;

[0008] Process the first feature map using the DCA (Discriminant Correlation Analysis) algorithm to obtain a first transformed feature map, and process the second feature map using the DCA algorithm to obtain a second transformed feature map;

[0009] Process the first transformed feature map using the SE-Net (Squeeze-and-Excitation Networks) algorithm, fuse the processing result with the second transformed feature map, and obtain an enhanced feature map based on the fusion result.

[0010] In some embodiments, the obtaining of the first feature map and the second feature map with the same size as the first feature map obtained after preprocessing the first feature map includes:

[0011] Obtain the first feature map;

[0012] Perform pooling and upsampling processing on the first feature map to obtain a second feature map with the same size as the first feature map.

[0013] In some embodiments, the obtaining of the enhanced feature map based on the fusion result includes:

[0014] Use a preset PCA (Principal Component Analysis) algorithm to reduce the dimension of the fusion result to obtain the enhanced feature map.

[0015] In some embodiments, the DCA algorithm includes row compression processing and column compression processing.

[0016] In a second aspect, a data processing method is provided in this embodiment, including:

[0017] Obtain target data; wherein, the target data is image data of a target object;

[0018] Use the principal component extraction algorithm in the trained network structure based on SE-Net and DCA to process the target data to obtain a first feature map;

[0019] Based on the network structure, perform enhancement processing on the first feature map according to the feature enhancement method described in the first aspect above to obtain an enhanced feature map;

[0020] Use the network structure to process the enhanced feature map to obtain the recognition result of the target object.

[0021] In some embodiments, the use of the principal component extraction algorithm in the trained network structure based on SE-Net and DCA to process the target data to obtain a first feature map includes:

[0022] Using the two-dimensional principal component extraction algorithm in the trained network structure based on SE-Net and DCA, perform row-column compression processing on the target data to obtain the principal component data of the target data;

[0023] Perform convolution processing on the principal component data to obtain the first feature map.

[0024] In some embodiments, before using the principal component extraction algorithm in the trained network structure based on SE-Net and DCA to process the target data to obtain the first feature map, the method further includes:

[0025] Obtain multi-pose training data;

[0026] Input the multi-pose training data into a preset network structure based on SE-Net and DCA for training to obtain the trained network structure based on SE-Net and DCA.

[0027] In some embodiments, the obtaining of the multi-pose training data includes:

[0028] Collect original training data;

[0029] Based on a preset three-dimensional multi-angle projection algorithm, process the original training data to obtain multi-pose training data.

[0030] In some embodiments, the processing of the original training data based on a preset three-dimensional multi-angle projection algorithm to obtain multi-pose training data includes:

[0031] Use a preset target detection algorithm to remove the image background information in the original training data to obtain first training data;

[0032] Use a preset three-dimensional reconstruction algorithm to perform three-dimensional reconstruction on the first training data to obtain a three-dimensional model;

[0033] Use a preset perspective projection transformation method to process the three-dimensional model to obtain the multi-pose training data.

[0034] In a third aspect, in this embodiment, a feature enhancement device for image recognition is provided, including: an acquisition module, a first processing module, and a second processing module, where:

[0035] The acquisition module is used to acquire a first feature map and a second feature map of the same size as the first feature map obtained after preprocessing the first feature map; wherein, the first feature map is a feature map obtained by performing feature extraction on a target image;

[0036] The first processing module is configured to process the first feature map using the DCA algorithm to obtain a first transformed feature map, and process the second feature map using the DCA algorithm to obtain a second transformed feature map;

[0037] The second processing module is configured to process the first transformed feature map using the SE-Net algorithm, fuse the processing result with the second transformed feature map, and obtain an enhanced feature map based on the fusion result.

[0038] In a fourth aspect, an electronic device is provided in this embodiment, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the feature enhancement method described in the first aspect above is implemented.

[0039] In a fifth aspect, a storage medium is provided in this embodiment, on which a computer program is stored. When the program is executed by a processor, the feature enhancement method described in the first aspect above is implemented.

[0040] This application provides a feature enhancement method, apparatus, data processing method, and electronic device. The feature enhancement method provided in this embodiment obtains a first feature map and a second feature map with the same size as the first feature map obtained after preprocessing the first feature map. The first feature map is a feature map obtained by performing feature extraction on a target image. The first feature map is processed using the DCA algorithm to obtain a first transformed feature map, and the second feature map is processed using the DCA algorithm to obtain a second transformed feature map. The first transformed feature map is processed using the SE-Net algorithm, the processing result is fused with the second transformed feature map, and an enhanced feature map is obtained based on the fusion result. It uses a network structure based on SE-Net and DCA to fuse image features according to the correlation information between pixel points and the correlation between structures in the dataset, thereby improving the accuracy of image recognition based on the fused features.

[0041] Details of one or more embodiments of this application are set forth in the following drawings and description to make other features, objects, and advantages of this application more concise and understandable. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The drawings described herein are used to provide a further understanding of this application, and form a part of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an improper limitation of this application. In the drawings:

[0043] Figure 1 is a hardware structure block diagram of a terminal of a feature enhancement method in the related art;

[0044] Figure 2 It is a flowchart of the feature enhancement method of this embodiment;

[0045] Figure 3 It is a structural diagram of the SE-DCA network of this embodiment;

[0046] Figure 4 It is a flowchart of the data processing method of this embodiment;

[0047] Figure 5 It is an overall network structure diagram of the network structure based on SE-Net and DCA of this embodiment;

[0048] Figure 6 It is a flowchart of the training method of the face recognition model of the preferred embodiment;

[0049] Figure 7 It is a structural block diagram of the data processing device of this embodiment. Detailed implementation manners

[0050] To understand the purpose, technical solution and advantages of this application more clearly, the following describes and explains this application in combination with the accompanying drawings and embodiments.

[0051] Unless otherwise defined, the technical terms or scientific terms involved in this application shall have the general meaning understood by those with ordinary skills in the technical field to which this application belongs. In this application, words such as "a", "one", "a kind of", "the", "these" and the like do not represent a limitation in quantity, and they can be singular or plural. The terms "including", "comprising", "having" and any variants thereof involved in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product or device including a series of steps or modules (units) is not limited to the listed steps or modules (units), but may include unlisted steps or modules (units), or may include other steps or modules (units) inherent in these processes, methods, products or devices. The terms "connected", "coupled" and the like involved in this application do not limit to physical or mechanical connections, but may include electrical connections, whether directly connected or indirectly connected. The term "plurality" involved in this application refers to two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" may represent: A exists alone, A and B exist simultaneously, and B exists alone. Usually, the character " / " represents an "or" relationship between the associated objects before and after. The terms "first", "second", "third" and the like involved in this application only distinguish similar objects and do not represent a specific sorting for the objects.

[0052] The method embodiments provided in this embodiment may be executed on a terminal, a computer, or a similar computing device. For example, when running on a terminal, Figure 1 is a hardware structural block diagram of the terminal of the feature enhancement method in this embodiment. As Figure 1 shown, the terminal may include one or more ( Figure 1 only one is shown in the figure) processors 102 and a memory 104 for storing data. Among them, the processor 102 may include, but is not limited to, a processing device such as a microprocessor MCU or a programmable logic device FPGA. The above terminal may further include a transmission device 106 for communication functions and an input / output device 108. Those of ordinary skill in the art can understand that Figure 1 the structure shown is only schematic and does not limit the structure of the above terminal. For example, the terminal may further include more or fewer components than Figure 1 shown in the figure, or have a different configuration from Figure 1 shown in the figure.

[0053] The memory 104 can be used to store computer programs. For example, software programs and modules of application software, such as the computer program corresponding to the feature enhancement method in this embodiment. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, that is, implements the above method. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memories, or other non-volatile solid-state memories. In some instances, the memory 104 may further include a memory remotely set relative to the processor 102, and these remote memories can be connected to the terminal through a network. Examples of the above network include, but are not limited to, the Internet, an enterprise intranet, a local area network, a mobile communication network, and combinations thereof.

[0054] The transmission device 106 is used to receive or send data via a network. The above network includes a wireless network provided by the communication provider of the terminal. In one instance, the transmission device 106 includes a network adapter (abbreviated as NIC), which can be connected to other network devices through a base station and thus can communicate with the Internet. In one instance, the transmission device 106 may be a radio frequency (abbreviated as RF) module, which is used to communicate with the Internet wirelessly.

[0055] In this embodiment, a feature enhancement method for image recognition is provided. Figure 2 is a flowchart of the feature enhancement method in this embodiment. As Figure 2 shown, the process includes the following steps:

[0056] Step S210: Obtain a first feature map and a second feature map of the same size as the first feature map obtained after preprocessing the first feature map; wherein, the first feature map is a feature map obtained by performing feature extraction on a target image.

[0057] Wherein, the target image can specifically be any image to be recognized, such as a face image, a human body image, a vehicle image, a license plate image, and an animal image, etc. Additionally, the (2D) 2 PCA (Two-directional two-dimensional PCA, two-dimensional principal component analysis method) algorithm can be used to perform row and column compression on the target image to obtain the principal component data of the target image, and perform convolution processing on the principal component data to obtain the feature map of the target image as the first feature map.

[0058] Additionally, after fusing the SE-Net algorithm and the DCA algorithm, the obtained SE-DCA network can be used to perform feature enhancement processing on the first feature map. Specifically, Figure 3 is the structural diagram of the SE-DCA network. As Figure 3 shown, after obtaining the first feature map, that is, Figure 3 Feature map1 in Figure 3 it, pooling (i.e., Figure 3 Max Pooling in Figure 3 it) and upsampling (i.e.,

[0059] Upsampling in

[0060] it) can be performed on the first feature map to obtain a second feature map of the same size as the first feature map (i.e., Figure 3 Feature map2 in Figure 3 it).

[0061] The DCA algorithm is used to achieve feature texture recognition and fusion, integrating the ideas of canonical correlation analysis and linear discriminant analysis. While ensuring the maximization of two types of feature data to be fused, it weakens the correlation between different classes in the data as much as possible. The algorithm principle of DCA specifically includes the following steps:

[0062] Step S220-1: Calculate the average value of each type of dataset in the dataset, as shown in the following formula.

[0063]

[0064] Among them, represents the mean of the i-th type of dataset, and n i represents the number of the i-th type of dataset, and x ij is the j-th data of the i-th type of dataset. Among them, in this embodiment, the category of the dataset can specifically be the category of the target object to which the data in the dataset belongs. For example, when the target object is a human face, the human faces of the same person are one category, and the human faces of different people belong to different categories.

[0065] Step S220-2, calculate the mean of the entire dataset, as shown in the following formula.

[0066]

[0067] Among them, n represents the number of the dataset, represents the mean of the entire dataset.

[0068] Step S220-3, calculate the between-class scatter matrix based on the mean of the entire dataset, as shown in the following formula.

[0069]

[0070]

[0071] Among them, is a diagonalization matrix and satisfies the following formula:

[0072]

[0073] Among them, P is an orthogonal eigenvector matrix, and τ is an orthogonal matrix of non-negative eigenvalues.

[0074] Step S220-4, perform dimensionality reduction on the feature matrix X, as shown in the following formula.

[0075]

[0076]

[0077] Among them, the size of the feature matrix is p×n, is a transfer matrix.

[0078] Step S220-5, based on the above steps, perform the same operation on the matrix Y to obtain Y'.

[0079] Step S220-6, construct a covariance matrix according to X' and Y', as shown in the following formula.

[0080] S' xy= X'Y' T (8)

[0081] Performing singular value decomposition on the above formula, the following formula is obtained:

[0082] S' xy = UΣV T (9)

[0083] where Σ = U T S' xy V, and the left and right singular vectors are: W cx = UΣ -1 / 2 , W cy = VΣ -1 / 2 .

[0084] Step S220-7, convert the feature matrix X′ to X * , and convert the feature matrix Y′ to Y * , as shown in the following formula.

[0085]

[0086]

[0087] where, when using the DCA algorithm to process the first feature map and the second feature map of this embodiment, the first feature map is the matrix X in the above steps, and the first converted feature map is X * ; the second feature map is the matrix Y in the above steps, and the second converted feature map is Y * .

[0088] Furthermore, this embodiment has improved the DCA algorithm. When the DCA algorithm performs compression according to matrix correlation for the second time, row compression is replaced by column compression. Compared with the current way of the DCA algorithm to repeatedly compress the matrix in the row direction, the improved DCA algorithm in this embodiment can avoid the problem of pixel relationship imbalance in the row and column directions and reduce the complexity of the matrix.

[0089] Step S230, use the SE-Net algorithm to process the first converted feature map, fuse the processing result with the second converted feature map, and obtain an enhanced feature map based on the fusion result.

[0090] Specifically, using the SE-Net algorithm to fuse the first converted feature map and the second converted feature map may specifically include performing pooling (corresponding to Figure 3 Global Pooling in it), two fully connected layers (corresponding to Figure 3 two FCs in it, where FC stands for Full Connection), and activation (corresponding to Figure 3After being processed by Sigmoid in , it is fused with the first transformed feature map (corresponding to Figure 3 Scale in ), and then fused with the second transformed feature map (corresponding to Figure 3 Fusion in ), and PCA dimensionality reduction is performed on the fusion result to obtain an enhanced feature map. Further, since the SE-Net algorithm is introduced, this algorithm will increase the computational complexity of feature enhancement. Therefore, by using the principal component extraction algorithm to compress the target data, the computational complexity in the subsequent processing can be reduced.

[0091] Among them, the processing process of the SE-Net algorithm for the feature map specifically includes: performing pooling, fully connected, fully connected, and activation on the feature map to generate a new feature map, and then fusing the previous feature map and the new feature map. It trains the model by learning feature weights according to the loss function, so that the effective feature maps have large weights and the ineffective or low-effect feature maps have small weights, thereby improving the performance of the model. By introducing the fusion of the SE-Net algorithm and the DCA algorithm, during the fusion of multiple feature maps, the correlation between the pixel points in the data set and the correlation between the structures in the data set can be considered as factors for fusion, thereby improving the expression effect of the fused features.

[0092] In addition, the SE-DCA network after the fusion of SE-Net and DCA in this embodiment can improve the convolutional neural network, thereby obtaining a network structure based on SE-Net and DCA. For example, the VGGNet16 (GG-Very-Deep-16 CNN, deep convolutional network) network can be improved by using the SE-Net algorithm and the DCA algorithm, and combined with (2D) 2 PCA algorithm to jointly form the EP-(2D) 2 PCANet network structure.

[0093] In the above steps S210 to S230, by obtaining the first feature map and the second feature map with the same size as the first feature map obtained after preprocessing the first feature map, where the first feature map is the feature map obtained by performing feature extraction on the target image, using the DCA algorithm to process the first feature map to obtain the first transformed feature map, using the DCA algorithm to process the second feature map to obtain the second transformed feature map, using the SE-Net algorithm to process the first transformed feature map, and fusing the processing result with the second transformed feature map, and obtaining an enhanced feature map based on the fusion result. It uses the network structure based on SE-Net and DCA to fuse the image features according to the correlation information between pixel points and the correlation between structures in the data set, thereby improving the accuracy of image recognition based on the fused features.

[0094] In one embodiment, based on the above step S210, obtaining a first feature map and a second feature map with the same size as the first feature map obtained after preprocessing the first feature map specifically includes the following steps:

[0095] Step S211, obtaining the first feature map;

[0096] Step S212, performing pooling and upsampling processing on the first feature map to obtain a second feature map with the same size as the first feature map.

[0097] In one embodiment, based on the above step S230, obtaining an enhanced feature map based on the fusion result specifically includes the following steps:

[0098] Step S231, using a preset PCA algorithm to perform dimensionality reduction on the fusion result to obtain an enhanced feature map.

[0099] In one embodiment, the DCA algorithm includes row compression processing and column compression processing. In this embodiment, the DCA algorithm is improved. When the DCA algorithm performs compression according to matrix correlation for the second time, row compression is replaced by column compression, which can avoid the problem of pixel relationship imbalance in the row and column directions and reduce the complexity of the matrix.

[0100] In this embodiment, a data processing method is provided. Figure 4 is a flowchart of the data processing method in this embodiment, as Figure 4 shown, and this process includes the following steps:

[0101] Step S410, obtaining target data; wherein, the target data is image data of a target object.

[0102] Among them, the target object can specifically be a face, a human body, a vehicle, a license plate, an animal, etc. to be recognized. The target object can include one or more types. For example, the target object can only be a face, or can include both a face and a vehicle. The determination of the target object can be set according to the requirements of the actual application scenario, and no specific limitation is made here. Exemplarily, when the target object is a face, the target data can be image data containing the face to be recognized collected by a camera device.

[0103] Step S420, using the principal component extraction algorithm in the trained network structure based on SE-Net and DCA to process the target data to obtain a first feature map.

[0104] Specifically, the network structure based on SE-Net and DCA can be a network structure obtained by improving a convolutional neural network using the SE-Net algorithm and the DCA algorithm. For example, the convolutional network obtained by improving the VGGNet16 (GG-Very-Deep-16 CNN, a deep convolutional network) network using the above-mentioned SE-Net algorithm and DCA algorithm.

[0105] Additionally, the (2D) 2 PCA algorithm can be used to perform row-column compression on the target data to obtain the principal component data of the target data, and convolutional processing is performed on the principal component data to obtain the feature map of the target data as the first feature map. In this embodiment, by fusing the SE-Net algorithm, the DCA algorithm, and the (2D) 2 PCA algorithm, an EP-(2D) 2 PCANet network structure is formed.

[0106] Step S430: Based on the above network structure, according to the feature enhancement method of the above embodiment, the first feature map is enhanced to obtain an enhanced feature map.

[0107] Based on the first feature map obtained in the above step S420, the feature enhancement part in the network structure based on SE-Net and DCA, that is, the above-mentioned SE-DCA network, is used for processing, so as to obtain an enhanced feature map.

[0108] Step S440: The above network structure is used to process the enhanced feature map to obtain the recognition result of the target object.

[0109] Specifically, other parts of the network structure based on SE-Net and DCA may further include several convolutional layers and pooling layers, and the processed data obtained after each layer of processing can be a new feature map. Additionally, the network structure is also provided with a fully connected layer for performing recognition processing on the data. By inputting the processed second feature map into other parts of the network structure based on SE-Net and DCA for processing, after multiple convolutional processing and pooling processing, the obtained processed data can be input into this fully connected layer. By using the fully connected layer to perform recognition processing on the processed data, the recognition result of the target object can be obtained. Corresponding to the type of the target object, the recognition result can specifically be the identity identifier of the target object and the attributes of the target object. For example, face attributes, pose categories, vehicle attributes, license plate numbers, and animal attributes, etc.

[0110] Based on the above content, it can be known that the overall network structure diagram of the network structure based on SE-Net and DCA provided in this embodiment is as shown in Figure 5As shown. Among them, the network structure 50 is the current VGGNet16. The network structure 501 therein can be replaced by the network structure 503 after the fusion of SE-Net and DCA on the right side of the figure, as shown by the bidirectional arrow in the figure, so as to realize the improvement of VGGNet16 and obtain the network structure based on SE-Net and DCA in this embodiment. Among them, in the improved network structure, (2D) 2 The PCA algorithm is used to extract the principal components of the target data to obtain the principal component data of the target data. The principal component data is subjected to convolution processing to obtain the first feature map of the target data. The above SE-DCA algorithm is used to perform feature enhancement processing on the first feature map to replace the pooling processing in the current recognition method, and an enhanced feature map is obtained. Then, the enhanced feature map is input into other network structures, such as other convolutional layers and pooling layers in the figure, for processing to obtain processed data, and finally, the fully connected layer, that is, the FC / 3+softmax network layer in the figure, is used to perform recognition processing on the processed data to obtain the recognition result of the target object. Additionally, the Figure 5 network structure 503 in can be used to replace the network structure 502 to realize the improvement of VGGNet16. In this embodiment, by introducing the fusion of the SE-Net algorithm and the DCA algorithm, it is possible to realize the enhancement processing of the feature map based on the correlation relationship between the pixel points in the data set and the correlation between the structures in the data set. Compared with the current feature map fusion method of directly adding the pixels at the corresponding positions on the feature map, this embodiment can enhance the expression effect of the features, thereby improving the accuracy of target object recognition.

[0111] Additionally, before using the trained network structure based on SE-Net and DCA to enhance the features of the target data, a multi-scale training data set can also be input into the preset network structure based on SE-Net and DCA for training, so as to obtain the trained network structure based on SE-Net and DCA. Specifically, a multi-pose training data is obtained by a preset three-dimensional multi-angle projection method, and the multi-pose training data is input into the preset network structure based on SE-Net and DCA for training, so as to reduce the impact of pose occlusion on the accuracy of target recognition.

[0112] In the above steps S410 to S440, by obtaining target data; wherein, the target data is image data of a target object, using the principal component extraction algorithm in the trained network structure based on SE-Net and DCA, processing the target data, based on the above network structure, according to the feature enhancement method of the above embodiment, enhancing the first feature map to obtain an enhanced feature map, and using the recognition structure in the network structure based on SE-Net and DCA to perform recognition processing on the enhanced feature map to obtain the recognition result of the target object. It uses the network structure based on SE-Net and DCA to fuse image features according to the correlation information between pixel points and the correlation between structures in the dataset, thereby improving the accuracy of face recognition.

[0113] In one embodiment, based on the above step S420, using the principal component extraction algorithm in the trained network structure based on SE-Net and DCA to process the target data to obtain the first feature map, which specifically includes the following steps:

[0114] Step S421, using the two-dimensional principal component extraction algorithm in the trained network structure based on SE-Net and DCA to perform row-column compression processing on the target data to obtain the principal component data of the target data.

[0115] Step S422, performing convolution processing on the principal component data to obtain the first feature map.

[0116] Based on the above steps S421 and S422, performing row-column compression on the target data and extracting the principal component of the target data can reduce the data volume of the target data, thereby reducing the subsequent computation amount and improving the algorithm operation efficiency.

[0117] In one embodiment, before using the principal component extraction algorithm in the trained network structure based on SE-Net and DCA to process the target data to obtain the first feature map, the above data processing method further includes the following steps:

[0118] Step S450, obtaining multi-pose training data.

[0119] Step S460, inputting the multi-pose training data into a preset network structure based on SE-Net and DCA for training to obtain a trained network structure based on SE-Net and DCA.

[0120] Further, in one embodiment, based on the above step S250, obtaining multi-pose training data specifically includes the following steps:

[0121] Step S451, collecting original training data.

[0122] Among them, the original training data can specifically be image data obtained by collecting images of objects for training. For example, face image data containing multiple human faces.

[0123] Step S452: Based on a preset three-dimensional multi-angle projection algorithm, process the original training data to obtain multi-pose training data.

[0124] Specifically, the preset three-dimensional multi-angle projection algorithm can be an algorithm for performing multi-pose projection processing on the original training data, and specifically can include an object detection algorithm and a perspective projection transformation method. The multi-pose training data can specifically be data including multiple different-angle images of the object for training.

[0125] Furthermore, in one embodiment, based on the above step S452, based on a preset three-dimensional multi-angle projection algorithm, process the original training data to obtain multi-pose training data, which specifically includes the following steps:

[0126] Step S4521: Use a preset object detection algorithm to remove the image background information in the original training data to obtain first training data.

[0127] Specifically, the MTCNN (Multi-task Cascaded Convolutional Networks) algorithm can be used to perform object detection on the original training data to remove the background information of the original training data.

[0128] Step S4522: Use a preset three-dimensional reconstruction algorithm to perform three-dimensional reconstruction on the first training data to obtain a three-dimensional model.

[0129] Specifically, based on the 3DMM (3D Morphable Models) algorithm, perform three-dimensional reconstruction on the first training data obtained in the above step S4521 to obtain a three-dimensional model.

[0130] Step S4523: Use a preset perspective projection transformation method to process the three-dimensional model to obtain multi-pose training data.

[0131] Specifically, a preset perspective projection transformation method can be used to project the three-dimensional model at different angles to obtain two-dimensional images in different poses, and use the two-dimensional images in different poses to expand the original training data to form multi-pose training data. Exemplarily, when the above data processing method is used for face recognition, the original training data is face training data, and a preset perspective projection transformation method can be used to perform multi-angle projection on the three-dimensional face model obtained based on the above step S4522 to obtain two-dimensional face images in different poses, and further obtain a multi-pose face training dataset.

[0132] By using the preset three-dimensional multi-angle projection algorithm to expand the training dataset, the above steps can extract pose-invariant features, thereby reducing the impact of pose occlusion on the recognition accuracy of the target object.

[0133] The following describes and illustrates this embodiment through preferred embodiments.

[0134] Figure 6 is a flowchart of the training method of the face recognition model of this preferred embodiment, as Figure 6 shown, the face recognition model training method includes the following steps:

[0135] Step S610, fuse the SE-Net algorithm and the DCA algorithm to obtain the SE-DCA network;

[0136] Step S620, use the SE-DCA network proposed in step S610 to enhance the VGGNet16 network to obtain EP-(2D) 2 PCANet;

[0137] Step S630, use the MTCNN algorithm to detect faces in the original training data and remove background information;

[0138] Step S640, use the 3DMM algorithm to perform three-dimensional reconstruction on the face image obtained by removing background information in step S630;

[0139] Step S650, use the perspective projection transformation method to project the three-dimensional face model obtained in step S640 at different angles to obtain two-dimensional face images in different poses, and use the two-dimensional face images in different poses to expand the original training data to obtain an expanded face dataset;

[0140] Step S660, input the expanded face dataset into the EP-(2D) 2 PCANet network for training to obtain a face recognition model.

[0141] In this embodiment, a feature enhancement device is also provided. This device is used to implement the above embodiments and preferred implementation manners, and those that have been described will not be repeated here. The following terms "module", "unit", "sub-unit", etc. can be a combination of software and / or hardware that can achieve a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, implementation in hardware, or a combination of software and hardware is also possible and contemplated.

[0142] Figure 7 is a structural block diagram of the feature enhancement device 70 of this embodiment, as Figure 7As shown in the figure, the feature enhancement device 70 includes: an acquisition module 72, a first processing module 74, and a second processing module 76, where:

[0143] The acquisition module 72 is configured to acquire a first feature map and a second feature map having the same size as the first feature map obtained after preprocessing the first feature map; wherein, the first feature map is a feature map obtained by performing feature extraction on a target image;

[0144] The first processing module 74 is configured to process the first feature map by using the DCA algorithm to obtain a first transformed feature map, and process the second feature map by using the DCA algorithm to obtain a second transformed feature map;

[0145] The second processing module 76 is configured to process the first transformed feature map by using the SE-Net algorithm, fuse the processing result with the second transformed feature map, and obtain the enhanced feature map based on the fusion result.

[0146] The above-mentioned feature enhancement device 70 acquires a first feature map and a second feature map having the same size as the first feature map obtained after preprocessing the first feature map, where the first feature map is a feature map obtained by performing feature extraction on a target image, processes the first feature map by using the DCA algorithm to obtain a first transformed feature map, processes the second feature map by using the DCA algorithm to obtain a second transformed feature map, processes the first transformed feature map by using the SE-Net algorithm, fuses the processing result with the second transformed feature map, and obtains the enhanced feature map based on the fusion result. It uses a network structure based on SE-Net and DCA to fuse image features according to the correlation information between pixel points and the correlation between structures in the dataset, thereby improving the accuracy of image recognition based on the fused features.

[0147] It should be noted that the above-mentioned each module can be a functional module or a program module, and can be implemented either by software or by hardware. For the modules implemented by hardware, the above-mentioned each module can be located in the same processor; or the above-mentioned each module can also be located in different processors in any combined form.

[0148] In this embodiment, an electronic device is further provided, which includes a memory and a processor. A computer program is stored in the memory, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0149] Optionally, the above-mentioned electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the above-mentioned processor, and the input / output device is connected to the above-mentioned processor.

[0150] Optionally, in this embodiment, the above-mentioned processor may be configured to execute the following steps through a computer program:

[0151] Obtain a first feature map and a second feature map of the same size as the first feature map obtained after preprocessing the first feature map; wherein, the first feature map is a feature map obtained by performing feature extraction on a target image;

[0152] Process the first feature map using the DCA algorithm to obtain a first transformed feature map, and process the second feature map using the DCA algorithm to obtain a second transformed feature map;

[0153] Process the first transformed feature map using the SE-Net algorithm, fuse the processing result with the second transformed feature map, and obtain the enhanced feature map based on the fusion result.

[0154] It should be noted that specific examples in this embodiment may refer to the examples described in the above-mentioned embodiment and optional implementation manners, and will not be elaborated herein.

[0155] In addition, in combination with the data processing method provided in the above-mentioned embodiment, a storage medium may also be provided in this embodiment to implement it. A computer program is stored on the storage medium; when the computer program is executed by the processor, any one of the feature enhancement methods in the above-mentioned embodiment is implemented.

[0156] It should be understood that the specific embodiments described here are only used to explain this application, rather than to limit it. According to the embodiments provided in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the protection scope of the present application.

[0157] Obviously, the drawings are only some examples or embodiments of the present application. For those of ordinary skill in the art, the present application can also be applied to other similar situations based on these drawings without creative efforts. In addition, it can be understood that although the work done during the development process here may be complex and time-consuming, for those of ordinary skill in the art, certain design, manufacturing, or production changes based on the technical content disclosed in the present application are only conventional technical means and should not be regarded as insufficient disclosure of the present application.

[0158] As used in this application, the term "embodiment" means that the specific features, structures or characteristics described in connection with an embodiment may be included in at least one embodiment of this application. The phrase appears in various positions in the specification and does not necessarily mean the same embodiment, nor does it mean that it is independent or alternative to other embodiments and mutually exclusive. Those of ordinary skill in the art can clearly or implicitly understand that the embodiments described in this application can be combined with other embodiments without conflict.

[0159] The above-described embodiments merely represent several implementation manners of this application. The description thereof is relatively specific and detailed, but should not be construed as a limitation on the scope of patent protection. It should be noted that for those of ordinary skill in the art, without departing from the concept of this application, several modifications and improvements can still be made, and these all fall within the protection scope of this application. Therefore, the protection scope of this application shall be subject to the appended claims.

Claims

1. A feature enhancement method for image recognition, characterized in that, Including: Obtain a first feature map; wherein, the first feature map is a feature map obtained by performing feature extraction on a target image; Perform pooling and upsampling processing on the first feature map to obtain a second feature map with the same size as the first feature map; Process the first feature map using the DCA algorithm to obtain a first transformed feature map, and process the second feature map using the DCA algorithm to obtain a second transformed feature map; Process the first transformed feature map using the SE-Net algorithm, fuse the processing result with the second transformed feature map, and obtain an enhanced feature map based on the fusion result.

2. The feature enhancement method according to claim 1, wherein The obtaining the enhanced feature map based on the fusion result includes: Use a preset PCA algorithm to perform dimensionality reduction on the fusion result to obtain the enhanced feature map.

3. The feature enhancement method according to any one of claims 1 to 2, characterized in that The DCA algorithm includes row compression processing and column compression processing.

4. A data processing method, characterized in that, Including: Obtain target data; wherein, the target data is image data of a target object; Use the principal component extraction algorithm in a trained network structure based on SE-Net and DCA to process the target data to obtain a first feature map; Based on the network structure, perform enhancement processing on the first feature map according to the feature enhancement method described in any one of claims 1 to 3 to obtain an enhanced feature map; Use the network structure to process the enhanced feature map to obtain the recognition result of the target object.

5. The data processing method according to claim 4, wherein The using the principal component extraction algorithm in a trained network structure based on SE-Net and DCA to process the target data to obtain a first feature map includes: Use the two-dimensional principal component extraction algorithm in a trained network structure based on SE-Net and DCA to perform row and column compression processing on the target data to obtain the principal component data of the target data; Perform convolution processing on the principal component data to obtain the first feature map.

6. The data processing method according to claim 4, wherein Before using the principal component extraction algorithm in a trained network structure based on SE-Net and DCA to process the target data to obtain a first feature map, the method further includes: Obtain multi-pose training data; Input the multi-pose training data into a preset network structure based on SE-Net and DCA for training to obtain the trained network structure based on SE-Net and DCA.

7. The data processing method according to claim 6, characterized in that The obtaining the multi-pose training data includes: Collect original training data; Based on a preset three-dimensional multi-angle projection algorithm, process the original training data to obtain multi-pose training data.

8. The data processing method according to claim 7, wherein The based on a preset three-dimensional multi-angle projection algorithm, processing the original training data to obtain multi-pose training data includes: Use a preset target detection algorithm to remove the image background information in the original training data to obtain first training data; Use a preset three-dimensional reconstruction algorithm to perform three-dimensional reconstruction on the first training data to obtain a three-dimensional model; Use a preset perspective projection transformation method to process the three-dimensional model to obtain the multi-pose training data.

9. An feature enhancement device for image recognition, characterized in that, Including: An acquisition module, a preprocessing module, a first processing module, and a second processing module, wherein: The acquisition module is configured to acquire a first feature map, where the first feature map is a feature map obtained by performing feature extraction on a target image; The preprocessing module is configured to perform pooling and upsampling on the first feature map to obtain a second feature map with the same size as the first feature map; The first processing module is configured to process the first feature map using the DCA algorithm to obtain a first transformed feature map, and process the second feature map using the DCA algorithm to obtain a second transformed feature map; The second processing module is configured to process the first transformed feature map using the SE-Net algorithm, fuse the processing result with the second transformed feature map, and obtain an enhanced feature map based on the fusion result.

10. An electronic device, comprising a memory and a processor, characterized in that, A computer program is stored in the memory, and the processor is configured to run the computer program to execute the feature enhancement method according to any one of claims 1 to 3.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, the steps of the feature enhancement method according to any one of claims 1 to 3 are implemented.

Images