Image recognition method, device, equipment, medium and computer product
Through the combination of wavelet transformation and image recognition neural network, the problem of multiplication noise influence in SAR image recognition is solved, the accuracy and robustness of image recognition are improved, and the efficient recognition of target objects in SAR images is achieved.
Patent Information
- Application Number
- CN202111669103.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-30
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2041-12-30
AI Technical Summary
The existing SAR image recognition methods have poor robustness and anti-interference ability, resulting in low accuracy of image recognition, especially when facing multiplicative noise.
Wavelet transformation processing is used to convert the synthetic aperture radar image into image transformation data and wavelet transformation coefficients, and the image is multiplied noise filtered by the image recognition neural network model trained to converge, combined with the convolutional neural network and the preset activation function. Feature extraction, identify the target object.
It improves the accuracy of SAR image recognition, enhances the processing ability of multiplicative noise, and improves the robustness and anti-interference ability of image recognition.
Smart Images

Figure CN114241204B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of image recognition technology, and in particular to an image recognition method, device, equipment, medium and computer product. Background Art
[0002] With the continuous development of science and technology, people pay more and more attention to indoor security. For example, financial companies attach great importance to the safe storage of their inherent assets and mortgage assets. Indoor target detection radar is the first choice for indoor target detection because of its all-round and all-weather detection.
[0003] The imaging basis of synthetic aperture radar (full name: Synthetic Aperture Radar, abbreviated as SAR) is the coherence principle, which easily leads to speckle noise in SAR images and affects the accuracy of image target detection.
[0004] At present, the main method for image recognition of SAR images is through the constant false alarm rate detection algorithm, which mainly uses the statistical distribution characteristics of image grayscale for detection, but has poor robustness and anti-interference capabilities, resulting in low image recognition accuracy. Summary of the invention
[0005] The present invention provides an image recognition method, device, equipment, medium and computer product, which are used to solve the problem of low image recognition accuracy of the current SAR image recognition method.
[0006] A first aspect of the present invention provides an image recognition method, comprising:
[0007] Acquire a synthetic aperture radar image to be identified;
[0008] Performing wavelet transform processing on the synthetic aperture radar image to generate corresponding image transform data and wavelet transform coefficients;
[0009] The image transformation data and the wavelet transformation coefficients are input into an image recognition neural network model trained to convergence to recognize the target object in the synthetic aperture radar image.
[0010] Furthermore, in the above method, the step of obtaining the synthetic aperture radar image to be identified includes:
[0011] Acquire an initial image generated by a synthetic aperture radar;
[0012] The initial image is subjected to radiometric calibration and normalization processing to generate a synthetic aperture radar image to be identified.
[0013] Further, in the method as described above, the image recognition neural network model trained to convergence comprises: an input layer, a convolutional neural network and an output layer;
[0014] The step of inputting the image transformation data and the wavelet transformation coefficients into an image recognition neural network model trained to convergence to identify the target object in the synthetic aperture radar image comprises:
[0015] Using an input layer to input the image transformation data and the wavelet transformation coefficients into the convolutional neural network;
[0016] Using the convolutional neural network to perform feature extraction and noise reduction on the image transformation data according to the wavelet transformation coefficients to generate a corresponding target feature map;
[0017] The target feature map is identified using the output layer to identify the target object.
[0018] Further, in the method as described above, the convolutional neural network includes a convolution layer, a preset activation function and a pooling layer;
[0019] The using the convolutional neural network to perform feature extraction and noise reduction on the image transformation data according to the wavelet transformation coefficients to generate a corresponding target feature map includes:
[0020] Using the convolutional layer to extract features from the image transformation data to generate corresponding image features;
[0021] Using the preset activation function to perform multiplicative noise filtering, inverse wavelet transform and nonlinear processing on the image features to generate corresponding nonlinear features;
[0022] The pooling layer is used to reduce the dimension of the nonlinear features to generate the target feature map.
[0023] Furthermore, in the method described above, the preset activation function is:
[0024]
[0025] in, is the preset activation sub-function, is the wavelet coefficient threshold;
[0026] The default activation function is:
[0027]
[0028] Where y is the wavelet coefficient threshold.
[0029] Furthermore, the method as described above, before inputting the image transformation data and the wavelet transformation coefficients into an image recognition neural network model trained to convergence to recognize the target object in the synthetic aperture radar image, further comprises:
[0030] Acquire training samples, wherein the training samples include: RGB image samples of labeled target objects; the RGB image samples are obtained by taking pictures with a camera;
[0031] Inputting the training sample into a preset residual network Resnet101 to train the preset residual network Resnet101;
[0032] Using the focal loss function and the regression loss function to determine whether the preset residual network Resnet101 model meets the preset convergence condition;
[0033] If the preset residual network Resnet101 model meets the convergence condition, the preset residual network Resnet101 model is used as a teacher network to train the preset image recognition neural network model;
[0034] A preset image recognition neural network model that meets the convergence condition is determined as the image recognition neural network model trained to convergence.
[0035] A second aspect of the present invention provides an image recognition device, comprising:
[0036] An acquisition module, used for acquiring a synthetic aperture radar image to be identified;
[0037] A generating module, used for performing wavelet transform processing on the synthetic aperture radar image to generate corresponding image transform data and wavelet transform coefficients;
[0038] The recognition module is used to input the image transformation data and the wavelet transformation coefficients into an image recognition neural network model trained to convergence, so as to recognize the target object in the synthetic aperture radar image.
[0039] Furthermore, in the above device, the acquisition module is specifically used for:
[0040] An initial image generated by a synthetic aperture radar is acquired; and a radiation calibration process and a normalization process are performed on the initial image to generate a synthetic aperture radar image to be identified.
[0041] Further, in the above-mentioned device, the image recognition neural network model trained to convergence includes: an input layer, a convolutional neural network and an output layer;
[0042] The identification module is specifically used for:
[0043] The image transformation data and the wavelet transformation coefficients are input into the convolutional neural network using an input layer; the convolutional neural network is used to perform feature extraction and noise reduction on the image transformation data according to the wavelet transformation coefficients to generate a corresponding target feature map; the target feature map is identified using an output layer to identify the target object.
[0044] Further, in the device as described above, the convolutional neural network includes a convolution layer, a preset activation function and a pooling layer;
[0045] When the recognition module uses the convolutional neural network to perform feature extraction and noise reduction on the image transformation data according to the wavelet transformation coefficients to generate a corresponding target feature map, it is specifically used to:
[0046] The convolution layer is used to extract features from the image transformation data to generate corresponding image features; the preset activation function is used to perform multiplicative noise filtering, inverse wavelet transform and nonlinear processing on the image features to generate corresponding nonlinear features; the pooling layer is used to perform dimensionality reduction processing on the nonlinear features to generate the target feature map.
[0047] Furthermore, in the above device, the preset activation function is:
[0048]
[0049] in, is the preset activation sub-function, is the wavelet coefficient threshold;
[0050] The default activation function is:
[0051]
[0052] Where y is the wavelet coefficient threshold.
[0053] Furthermore, the device as described above further comprises:
[0054] A training module is used to obtain training samples, wherein the training samples include: RGB image samples of labeled target objects; the RGB image samples are obtained by taking pictures with a camera; the training samples are input into a preset residual network Resnet101 to train the preset residual network Resnet101; a focal loss loss function and a regression loss function are used to determine whether the preset residual network Resnet101 model meets the preset convergence conditions; if the preset residual network Resnet101 model meets the convergence conditions, the preset residual network Resnet101 model is used as a teacher network to train the preset image recognition neural network model; the preset image recognition neural network model that meets the convergence conditions is determined as the image recognition neural network model trained to convergence.
[0055] A third aspect of the present invention provides an electronic device, comprising: a processor, and a memory communicatively connected to the processor;
[0056] The memory stores computer-executable instructions;
[0057] The processor executes the computer-executable instructions stored in the memory to implement the image recognition method as described in any one of the first aspects.
[0058] A fourth aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by a processor, they are used to implement the image recognition method described in any one of the first aspects.
[0059] A fifth aspect of the present invention provides a computer program product, comprising a computer program, which, when executed by a processor, implements the image recognition method described in any one of the first aspects.
[0060] The present invention provides an image recognition method, device, equipment, medium and computer product, the method comprising: obtaining a synthetic aperture radar image to be recognized; performing wavelet transform processing on the synthetic aperture radar image to generate corresponding image transformation data and wavelet transform coefficients; inputting the image transformation data and the wavelet transform coefficients into an image recognition neural network model trained to convergence to recognize the target object in the synthetic aperture radar image. The image recognition method of the present invention converts the synthetic aperture radar image into image transformation data and wavelet transform coefficients through wavelet transform processing, and then performs multiplicative noise filtering and noise reduction on the synthetic aperture radar image based on wavelet transform through the image recognition neural network model trained to convergence, and recognizes the target object in the image, thereby improving the accuracy of image recognition. BRIEF DESCRIPTION OF THE DRAWINGS
[0061] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.
[0062] Figure 1 A scene graph that can implement the image recognition method of the embodiment of the present invention;
[0063] Figure 2 A schematic diagram of a flow chart of an image recognition method provided by a first embodiment of the present invention;
[0064] Figure 3 A schematic diagram of a flow chart of an image recognition method provided by a second embodiment of the present invention;
[0065] Figure 4 A schematic diagram of the structure of an image recognition device provided by a third embodiment of the present invention;
[0066] Figure 5 A schematic diagram of the structure of an image recognition device provided by a fourth embodiment of the present invention;
[0067] Figure 6 A schematic structural diagram of an electronic device provided in a fifth embodiment of the present invention.
[0068] The above drawings have shown clear embodiments of the present invention, which will be described in more detail below. These drawings and text descriptions are not intended to limit the scope of the present invention in any way, but to illustrate the concept of the present invention for those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION
[0069] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are merely examples of devices and methods consistent with some aspects of the present invention as detailed in the appended claims.
[0070] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present invention will be described below in conjunction with the accompanying drawings.
[0071] In order to clearly understand the technical solution of the present application, the solution of the prior art is first introduced in detail. The imaging basis of synthetic aperture radar (full name in English: Synthetic Aperture Radar, English abbreviation: SAR) is the coherence principle. In the radar echo signal, the grayscale values of adjacent pixels will produce some random changes due to coherence, and this random change is carried out around a certain mean, thus generating speckle noise in the image. The generation of speckle noise is a defect caused by the coherence principle on which SAR imaging is based, so it is inevitable. In a uniform area, the more severely contaminated the area is, the brighter it appears in the image.
[0072] Multiplicative noise has a great influence on SAR images. At present, the image recognition method of SAR images is mainly through the constant false alarm rate detection algorithm, which mainly uses the statistical distribution characteristics of image grayscale for detection. It has poor robustness and anti-interference ability and cannot solve the influence of multiplicative noise, resulting in low image recognition accuracy.
[0073] Therefore, in order to solve the problem of low image recognition accuracy of the image recognition method of SAR images in the prior art, the inventors found in the research that in order to solve this problem, multiplicative noise filtering can be performed through wavelet transform, thereby solving the problem of multiplicative noise in SAR images. At the same time, the image recognition network model can be optimized, and some algorithms in the image recognition network model can be combined with wavelet transform, such as fusing the activation function with the wavelet transform coefficients.
[0074] Specifically, firstly, a synthetic aperture radar image to be identified is obtained, and the synthetic aperture radar image is subjected to wavelet transform processing to generate corresponding image transform data and wavelet transform coefficients. Then, the image transform data and wavelet transform coefficients are input into an image recognition neural network model trained to convergence to identify the target object in the synthetic aperture radar image. Based on the image recognition neural network model trained to convergence and wavelet transform, the synthetic aperture radar image is subjected to multiplicative noise filtering and denoising, and the target object in the image is identified, thereby improving the accuracy of image recognition.
[0075] Based on the above creative findings, the inventor proposed the technical solution of the present application.
[0076] The following is an introduction to the application scenarios of the image recognition method provided by the embodiment of the present invention. Figure 1As shown, 1 is a first electronic device and 2 is a second electronic device. The second electronic device 2 stores a synthetic aperture radar image to be identified. When image recognition is required, the first electronic device 1 can obtain the synthetic aperture radar image to be identified from the second electronic device 2. Then, the first electronic device 1 performs wavelet transform processing on the synthetic aperture radar image to generate corresponding image transformation data and wavelet transform coefficients. The first electronic device 1 inputs the image transformation data and the wavelet transform coefficients into an image recognition neural network model trained to convergence to identify the target object in the synthetic aperture radar image. In this application scenario, the image recognition neural network model trained to convergence can be based on the yolo_v1 network model, and the activation function therein is fused with the wavelet transform coefficients. Thereby realizing the function of multiplicative noise filtering, and not affecting the original recognition function of the image recognition neural network model, thereby improving the accuracy of image recognition. At the same time, the identified target object can be output to a display device for user viewing, or output to other electronic devices for subsequent analysis and processing.
[0077] The embodiments of the present invention are described below in conjunction with the accompanying drawings.
[0078] Figure 2 A schematic diagram of a flow chart of an image recognition method provided by the first embodiment of the present invention is shown in FIG. Figure 2 As shown, in this embodiment, the execution subject of the embodiment of the present invention is an image recognition device, which can be integrated into an electronic device. The image recognition method provided by this embodiment includes the following steps:
[0079] Step S101, obtaining a synthetic aperture radar image to be identified.
[0080] In this embodiment, the synthetic aperture radar image, that is, the SAR image, is obtained after preprocessing of the image generated by the synthetic aperture radar. The preprocessing may include radiation calibration processing, that is, converting the brightness grayscale value of the image into absolute radiation brightness, normalization processing, etc.
[0081] The synthetic aperture radar image to be identified can be stored in a database after preprocessing, and can be obtained when the synthetic aperture radar image needs to be identified.
[0082] Step S102: performing wavelet transform processing on the synthetic aperture radar image to generate corresponding image transform data and wavelet transform coefficients.
[0083] In this embodiment, the main processing process of the wavelet denoising algorithm is: performing wavelet decomposition on the noisy image to obtain wavelet transform coefficients. Performing corresponding threshold processing on the wavelet transform coefficients to eliminate noise as much as possible and retain image detail information. Performing inverse wavelet transform and reconstructing to obtain a denoised image.
[0084] Therefore, it is necessary to perform wavelet transform processing on the synthetic aperture radar image to determine the corresponding transformed image transformation data and wavelet transform coefficients.
[0085] Step S103: input the image transformation data and the wavelet transformation coefficients into an image recognition neural network model trained to convergence, so as to recognize the target object in the synthetic aperture radar image.
[0086] In this embodiment, the image recognition neural network model trained to convergence includes the process of performing corresponding threshold processing on wavelet transform coefficients and inverse wavelet transform, and also includes the process of feature extraction and object recognition on the image.
[0087] An embodiment of the present invention provides an image recognition method, the method comprising: obtaining a synthetic aperture radar image to be recognized, performing wavelet transform processing on the synthetic aperture radar image to generate corresponding image transformation data and wavelet transform coefficients, and inputting the image transformation data and the wavelet transform coefficients into an image recognition neural network model trained to convergence to recognize a target object in the synthetic aperture radar image.
[0088] The image recognition method of the present invention converts a synthetic aperture radar image into image transformation data and wavelet transformation coefficients through wavelet transformation processing, and then performs multiplicative noise filtering and noise reduction on the synthetic aperture radar image based on wavelet transformation through an image recognition neural network model trained to convergence, and identifies the target object in the image, thereby improving the accuracy of image recognition.
[0089] Figure 3 A flow chart of an image recognition method provided by the second embodiment of the present invention is shown in FIG. Figure 3 As shown, the image recognition method provided in this embodiment is based on the image recognition method provided in the previous embodiment of the present invention, and each step is further refined. The image recognition method provided in this embodiment includes the following steps.
[0090] It should be noted that the image recognition neural network model trained to convergence includes: an input layer, a convolutional neural network and an output layer.
[0091] Step S201, obtaining an initial image generated by a synthetic aperture radar.
[0092] In this embodiment, the synthetic aperture radar is generally configured to detect indoor objects, and the synthetic aperture radar generally detects and generates an initial image in real time.
[0093] Optionally, in this embodiment, before inputting the image transformation data and the wavelet transformation coefficients into the image recognition neural network model trained to convergence to recognize the target object in the synthetic aperture radar image, the method further includes:
[0094] Obtain training samples, including: RGB image samples of labeled target objects. The RGB image samples are obtained by taking pictures with a camera.
[0095] The training samples are input into the preset residual network Resnet101 to train the preset residual network Resnet101.
[0096] The focal loss function and regression loss function are used to determine whether the preset residual network Resnet101 model meets the preset convergence conditions.
[0097] If the preset residual network Resnet101 model meets the convergence conditions, the preset residual network Resnet101 model is used as the teacher network to train the preset image recognition neural network model.
[0098] A preset image recognition neural network model that meets the convergence condition is determined as the image recognition neural network model trained to convergence.
[0099] In this embodiment, the image recognition neural network model can be trained in advance. The training method adopted in this embodiment is to train the student network with the teacher network. The image recognition neural network model is preset as the student network, and the residual network Resnet101 model trained to convergence is used as the teacher network. The residual network Resnet101 model is a convolutional neural network with the advantage of being easy to optimize.
[0100] When training the residual network Resnet101 model and the image recognition neural network model, multiple frames of SAR images and RGB images can be obtained from a general product that is a fusion of indoor cameras and millimeter-wave radars. RGB is a color system.
[0101] Then, by manually annotating the RGB image, the corresponding target object in the RGB image is marked with a box.
[0102] Next, the 50,000 labeled RGB images are used as the input of the residual network Resnet101 model, and the focal loss loss function and regression loss function are used as loss functions to train the residual network Resnet101 model. Finally, the training is completed to obtain the teacher model needed later.
[0103] The image recognition neural network model, which serves as a student model, only needs to be continuously trained by the teacher model and can be used in the future.
[0104] In the present embodiment, the image recognition neural network model can be based on shufflenetV2 network and yolo_v1 network architecture, shufflenetV2 network is a kind of convolutional neural network, yolo_v1 network is a kind of target detection network, and also is a kind of object recognition and positioning algorithm based on deep neural network, and its biggest feature is that it runs very fast. Based on shufflenetV2 network and yolo_v1 network architecture, shufflenetV2 network can be used as backbone network, and yolo_v1 network can be used as recognition network for fusion, so as to further improve object recognition efficiency.
[0105] At the same time, it can also be based only on the yolo_v1 network, which is not limited in this embodiment.
[0106] Since the image recognition neural network model needs to integrate the wavelet transform capability, the relu activation layer in the image recognition neural network model can be integrated with the wavelet transform. For example, in the method of adding nonlinear features to the relu activation layer algorithm, a layer of wavelet transform coefficient determination is added. This fusion method can be integrated only with the rulu1 activation layer, that is, in the first activation layer, and the normal neural network structure is maintained later. This enables the image recognition neural network model to have the function of filtering multiplicative noise.
[0107] At the same time, after constructing the teacher network and the student network, a knowledge distillation architecture can be formed. By introducing soft targets related to the teacher network as part of the total loss, the training of the student network can be induced to achieve knowledge transfer.
[0108] Step S202: performing radiometric calibration and normalization processing on the initial image to generate a synthetic aperture radar image to be identified.
[0109] In this embodiment, the preprocessing performed on the initial image includes radiation calibration processing and normalization processing, so as to preprocess the brightness grayscale value of the image.
[0110] Step S203: performing wavelet transform processing on the synthetic aperture radar image to generate corresponding image transformation data and wavelet transform coefficients.
[0111] In this embodiment, the implementation of step 203 is similar to the implementation of step 102 in the previous embodiment of the present invention, and will not be described in detail here.
[0112] Step S204: Use the input layer to input the image transformation data and wavelet transformation coefficients into the convolutional neural network.
[0113] In this embodiment, the image recognition neural network model may include an input layer, a convolutional neural network and an output layer. The input layer generally has certain restrictions on the input image size. The convolutional neural network generally includes a convolutional layer, an activation layer and a pooling layer. The input layer generally has functional algorithms for object recognition and object position determination.
[0114] Step S205, using a convolutional neural network to perform feature extraction and noise reduction on the image transformation data according to the wavelet transformation coefficients to generate a corresponding target feature map.
[0115] In this embodiment, since the activation layer in the convolutional neural network is optimized, the image transformation data can be denoised according to the wavelet transformation coefficients.
[0116] Optionally, in this embodiment, the convolutional neural network includes a convolution layer, a preset activation function and a pooling layer.
[0117] The convolutional neural network is used to extract features and reduce noise from the image transformation data according to the wavelet transform coefficients to generate the corresponding target feature map, including:
[0118] The convolution layer is used to extract features from the image transformation data to generate corresponding image features.
[0119] The preset activation function is used to perform multiplicative noise filtering, inverse wavelet transform and nonlinear processing on the image features to generate corresponding nonlinear features.
[0120] The pooling layer is used to reduce the dimension of nonlinear features to generate the target feature map.
[0121] In this embodiment, after the multiplicative noise filtering process, the processed image data needs to be subjected to inverse wavelet transform processing, thereby completing the entire process of wavelet denoising. The main function of the pooling layer is to perform dimensionality reduction processing. Since the image data initially input is generally large, such as a size of 448*448, it needs to be continuously reduced to a similar size of 49*49 through the pooling layer.
[0122] Optionally, in this embodiment, the preset activation function is:
[0123]
[0124] in, is the preset activation sub-function, is the wavelet coefficient threshold.
[0125] The default activation function is:
[0126]
[0127] Where y is the wavelet coefficient threshold.
[0128] In this embodiment, the filtering process can be seen from the preset activation function, that is, firstly, the value of g(x,y) is determined by the magnitude relationship between the wavelet transform coefficient threshold and the wavelet transform coefficient of the image transform data. If g(x,y) is 0, it means that it needs to be filtered out. After the wavelet transform filtering process, the nonlinear feature processing is performed through the determination of the original activation function.
[0129] Step S206: Use the output layer to identify the target feature map to identify the target object.
[0130] In this embodiment, the output layer is generally provided with an object recognition algorithm, such as the probability of recognizing an object, the position of an object, etc.
[0131] In the image recognition method of this embodiment, the main body of the network adopts a single-stage detector such as the yolo_v1 network model, which has a light structure. By combining wavelet transform, the feature dimension is reduced without introducing prior knowledge, the amount of data is reduced, the network efficiency is improved, and the hardware occupation is reduced. At the same time, multiplication filtering can be implemented in a convolutional neural network (full name: Convolutional Neural Networks, abbreviated as: CNN), and the detection accuracy on SAR images is high.
[0132] Figure 4 A schematic diagram of the structure of an image recognition device provided in the third embodiment of the present invention is shown in FIG. Figure 4 As shown, in this embodiment, the image recognition device 300 includes:
[0133] The acquisition module 301 is used to acquire a synthetic aperture radar image to be identified.
[0134] The generating module 302 is used to perform wavelet transform processing on the synthetic aperture radar image to generate corresponding image transform data and wavelet transform coefficients.
[0135] The recognition module 303 is used to input the image transformation data and the wavelet transformation coefficients into the image recognition neural network model trained to convergence, so as to recognize the target object in the synthetic aperture radar image.
[0136] The image recognition device provided in this embodiment can perform Figure 2 The technical solution of the method embodiment shown in the figure has the same implementation principle and technical effect as Figure 2 The method embodiments shown are similar and will not be described in detail here.
[0137] at the same time, Figure 5 A schematic diagram of the structure of an image recognition device provided in a fourth embodiment of the present invention is shown in FIG. Figure 5As shown, in order to better distinguish the image recognition device of this embodiment from the previous embodiment, the image recognition device of the previous embodiment is the image recognition device 300, and the image recognition device provided by the present invention is the image recognition device 400. Both refer to image recognition devices, but the modules included are not exactly the same.
[0138] The image recognition device 400 provided by the present invention is further provided with a training module 401 .
[0139] Optionally, in this embodiment, the acquisition module 301 is specifically used for:
[0140] An initial image generated by a synthetic aperture radar is obtained, and a radiometric calibration process and a normalization process are performed on the initial image to generate a synthetic aperture radar image to be identified.
[0141] Optionally, in this embodiment, the image recognition neural network model trained to convergence includes: an input layer, a convolutional neural network, and an output layer.
[0142] The identification module 303 is specifically used for:
[0143] The input layer is used to input the image transformation data and wavelet transformation coefficients into the convolutional neural network. The convolutional neural network is used to extract features and reduce noise on the image transformation data according to the wavelet transformation coefficients to generate the corresponding target feature map. The output layer is used to identify the target feature map to identify the target object.
[0144] Optionally, in this embodiment, the convolutional neural network includes a convolution layer, a preset activation function and a pooling layer.
[0145] When the recognition module 303 uses a convolutional neural network to perform feature extraction and noise reduction on the image transformation data according to the wavelet transformation coefficients to generate a corresponding target feature map, it is specifically used to:
[0146] The convolution layer is used to extract features from the image transformation data to generate the corresponding image features. The preset activation function is used to perform multiplicative noise filtering, inverse wavelet transformation and nonlinear processing on the image features to generate the corresponding nonlinear features. The pooling layer is used to reduce the dimension of the nonlinear features to generate the target feature map.
[0147] Optionally, in this embodiment, the preset activation function is:
[0148]
[0149] in, is the preset activation sub-function, is the wavelet coefficient threshold.
[0150] The default activation function is:
[0151]
[0152] Where y is the wavelet coefficient threshold.
[0153] Optionally, in this embodiment, the image recognition device 400 further includes:
[0154] The training module 401 is used to obtain training samples, and the training samples include: RGB image samples of the labeled target object. The RGB image samples are obtained by taking pictures with a camera. The training samples are input into the preset residual network Resnet101 to train the preset residual network Resnet101. The focal loss loss function and the regression loss function are used to determine whether the preset residual network Resnet101 model meets the preset convergence conditions. If the preset residual network Resnet101 model meets the convergence conditions, the preset residual network Resnet101 model is used as a teacher network to train the preset image recognition neural network model. The preset image recognition neural network model that meets the convergence conditions is determined as the image recognition neural network model trained to convergence.
[0155] The image recognition device provided in this embodiment can perform Figure 2-Figure 3 The technical solution of the method embodiment shown in the figure has the same implementation principle and technical effect as Figure 2-Figure 3 The method embodiments shown are similar and will not be described in detail here.
[0156] According to an embodiment of the present invention, the present invention also provides an electronic device, a computer-readable storage medium and a computer program product.
[0157] like Figure 6 As shown, Figure 6 1 is a schematic diagram of the structure of an electronic device provided by the fifth embodiment of the present invention. The electronic device is intended to be various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present invention described and / or claimed herein.
[0158] like Figure 6 As shown, the electronic device includes: a processor 501 and a memory 502. The various components are connected to each other using different buses and can be installed on a common mainboard or in other ways as required. The processor can process instructions executed in the electronic device.
[0159] The memory 502 is a non-transitory computer-readable storage medium provided by the present invention. The memory stores instructions executable by at least one processor to enable at least one processor to perform the image recognition method provided by the present invention. The non-transitory computer-readable storage medium of the present invention stores computer instructions, which are used to enable a computer to perform the image recognition method provided by the present invention.
[0160] The memory 502 is a non-transient computer-readable storage medium that can be used to store non-transient software programs, non-transient computer executable programs and modules, such as program instructions / modules corresponding to the image recognition method in the embodiment of the present invention (for example, the attached Figure 4 The processor 501 executes various functional applications and data processing of the server by running the non-transient software programs, instructions and modules stored in the memory 502, that is, the image recognition method in the above method embodiment is implemented.
[0161] At the same time, this embodiment also provides a computer product. When the instructions in the computer product are executed by a processor of an electronic device, the electronic device can execute the image recognition methods of the above-mentioned embodiments one to two.
[0162] Those skilled in the art will readily appreciate other implementations of the embodiments of the present invention after considering the specification and practicing the invention disclosed herein. The present invention is intended to cover any variations, uses, or adaptations of the embodiments of the present invention that follow the general principles of the embodiments of the present invention and include common knowledge or customary technical means in the art that are not disclosed in the embodiments of the present invention. The specification and examples are to be regarded as exemplary only, and the true scope and spirit of the embodiments of the present invention are indicated by the claims.
[0163] It should be understood that the embodiments of the present invention are not limited to the precise structures described above and shown in the drawings, and various modifications and changes may be made without departing from the scope thereof. The scope of the embodiments of the present invention is limited only by the appended claims.
Claims
1. An image recognition method, It is characterized in that include: Acquire a synthetic aperture radar image to be identified; Performing wavelet transform processing on the synthetic aperture radar image to generate corresponding image transform data and wavelet transform coefficients; Inputting the image transformation data and the wavelet transformation coefficients into an image recognition neural network model trained to convergence to recognize a target object in the synthetic aperture radar image; The step of obtaining a synthetic aperture radar image to be identified comprises: Acquire an initial image generated by a synthetic aperture radar; Performing radiometric calibration and normalization processing on the initial image to generate a synthetic aperture radar image to be identified; The image recognition neural network model trained to convergence includes: an input layer, a convolutional neural network and an output layer; The step of inputting the image transformation data and the wavelet transformation coefficients into an image recognition neural network model trained to convergence to identify the target object in the synthetic aperture radar image comprises: Using an input layer to input the image transformation data and the wavelet transformation coefficients into the convolutional neural network; Using the convolutional neural network to perform feature extraction and noise reduction on the image transformation data according to the wavelet transformation coefficients to generate a corresponding target feature map; Using the output layer to identify the target feature map to identify the target object; The convolutional neural network includes a convolution layer, a preset activation function and a pooling layer; The using the convolutional neural network to perform feature extraction and noise reduction on the image transformation data according to the wavelet transformation coefficients to generate a corresponding target feature map includes: Using the convolutional layer to extract features from the image transformation data to generate corresponding image features; Using the preset activation function to perform multiplicative noise filtering, inverse wavelet transform and nonlinear processing on the image features to generate corresponding nonlinear features; The pooling layer is used to reduce the dimension of the nonlinear features to generate the target feature map; The preset activation function is: in, is the preset activation sub-function, is the wavelet coefficient threshold; The default activation function is: Where y is the wavelet coefficient threshold.
2. The method according to claim 1, It is characterized in that Before inputting the image transformation data and the wavelet transformation coefficients into an image recognition neural network model trained to convergence to recognize the target object in the synthetic aperture radar image, the method further includes: Acquire training samples, wherein the training samples include: RGB image samples of labeled target objects; the RGB image samples are obtained by taking pictures with a camera; Inputting the training sample into a preset residual network Resnet101 to train the preset residual network Resnet101; Using the focal loss function and the regression loss function to determine whether the preset residual network Resnet101 model meets the preset convergence condition; If the preset residual network Resnet101 model meets the convergence condition, the preset residual network Resnet101 model is used as a teacher network to train the preset image recognition neural network model; A preset image recognition neural network model that meets the convergence condition is determined as the image recognition neural network model trained to convergence.
3. An image recognition device, It is characterized in that include: An acquisition module, used for acquiring a synthetic aperture radar image to be identified; A generating module, used for performing wavelet transform processing on the synthetic aperture radar image to generate corresponding image transform data and wavelet transform coefficients; A recognition module, configured to input the image transformation data and the wavelet transformation coefficients into an image recognition neural network model trained to convergence, so as to recognize a target object in the synthetic aperture radar image; The acquisition module is specifically used to acquire an initial image generated by a synthetic aperture radar; Performing radiometric calibration and normalization processing on the initial image to generate a synthetic aperture radar image to be identified; The image recognition neural network model trained to convergence includes: an input layer, a convolutional neural network and an output layer; The recognition module is specifically used to input the image transformation data and the wavelet transformation coefficient into the convolutional neural network using an input layer; perform feature extraction and noise reduction processing on the image transformation data according to the wavelet transformation coefficient to generate a corresponding target feature map; and recognize the target feature map using an output layer to recognize the target object; the convolutional neural network includes a convolutional layer, a preset activation function and a pooling layer; The recognition module, when using the convolutional neural network to perform feature extraction and noise reduction on the image transformation data according to the wavelet transformation coefficients to generate a corresponding target feature map, is specifically used to use the convolution layer to perform feature extraction on the image transformation data to generate corresponding image features; use the preset activation function to perform multiplicative noise filtering, inverse wavelet transformation and nonlinear processing on the image features to generate corresponding nonlinear features; use the pooling layer to perform dimensionality reduction processing on the nonlinear features to generate the target feature map; The preset activation function is: in, is the preset activation sub-function, is the wavelet coefficient threshold; The default activation function is: Where y is the wavelet coefficient threshold.
4. An electronic device, It is characterized in that include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executable instructions; The processor executes the computer-executable instructions stored in the memory to implement the image recognition method according to any one of claims 1 to 2.
5. A computer-readable storage medium, It is characterized in that The computer-readable storage medium stores computer-executable instructions, which are used to implement the image recognition method according to any one of claims 1 to 2 when executed by a processor.
6. A computer program product comprising a computer program, It is characterized in that When the computer program is executed by a processor, the image recognition method according to any one of claims 1 to 2 is implemented.
