Class estimation method and apparatus, electronic device, and storage medium

Abstract

The application provides a category estimation method, device, electronic equipment and storage medium. The category estimation method comprises: obtaining an initial image; the initial image comprises a target object to be processed; performing feature extraction on the initial image to obtain a spatial feature image; processing the spatial feature image to obtain a frequency domain feature image; performing feature filtering on the frequency domain feature image to obtain a target feature image; performing dimension reduction processing on the target feature image to obtain an intermediate feature image; and obtaining a category estimation of the target object based on the intermediate feature image. The method can improve the category estimation accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and in particular to a category estimation method and device, electronic equipment and storage medium. BACKGROUND

[0002] With the development of image processing technology, it is widely used in various fields, for example, image classification or image object detection. Among them, image object detection is to locate the target of interest from the image by using the theory and method of image processing and pattern recognition, which needs to accurately judge the specific category of each target. Image object detection has wide application in face recognition, medical imaging, intelligent video monitoring, robot navigation and augmented reality.

[0003] In the prior art, an unsupervised domain adaptation method is generally used for image object detection, which often has difficulty in aligning data distribution and has the problem of low category estimation accuracy. SUMMARY

[0004] The present application provides a category estimation method and device, electronic equipment and storage medium to solve the defects of poor category estimation effect in the prior art, and achieve the purpose of improving the category estimation accuracy.

[0005] The present application provides a category estimation method, comprising: obtaining an initial image; the initial image includes a target object to be processed; performing feature extraction on the initial image to obtain a spatial feature image; processing the spatial feature image to obtain a frequency domain feature image; performing feature filtering on the frequency domain feature image to obtain a target feature image; performing dimension reduction processing on the target feature image to obtain an intermediate feature image; and obtaining a category estimation of the target object based on the intermediate feature image.

[0006] According to the category estimation method provided by the present application, the frequency domain feature image includes first frequency domain features and second frequency domain features; and the feature filtering on the frequency domain feature image to obtain the target feature image comprises: filtering the first frequency domain features in the frequency domain feature image according to a frequency threshold to obtain the target feature image; wherein the target feature image includes the second frequency domain features.

[0007] According to the category estimation method provided by the present application, the filtering of the first frequency domain features in the frequency domain feature image to obtain the target feature image comprises: obtaining a frequency domain filtering function; converting the frequency domain filtering function from the frequency domain to the spatial domain to obtain a spatial domain filtering function; and filtering the first frequency domain features in the frequency domain feature image by using the spatial domain filtering function to obtain the target feature image.

[0008] According to a category estimation method provided by the present invention, processing the spatial feature image to obtain a frequency domain feature image includes: performing a discrete Fourier transform on the spatial feature image to obtain the frequency domain feature image; and performing feature filtering on the frequency domain feature image to obtain a target feature image includes: performing feature filtering on the frequency domain feature image to obtain a filtered feature image; and performing an inverse discrete Fourier transform on the filtered feature image to obtain the target feature image.

[0009] According to a category estimation method provided by the present invention, the step of extracting features from the initial image to obtain a spatial feature image includes: extracting features from the initial image using a convolutional layer of a neural network to obtain a spatial feature image.

[0010] According to a category estimation method provided by the present invention, the convolutional layer includes a strided convolutional layer; the step of using the convolutional layer of the neural network to extract features from the initial image to obtain a spatial feature image includes: replacing the strided convolutional layer with a frequency module to extract features from the initial image to obtain a spatial feature image; wherein, the frequency module is used to perform a step of processing the spatial feature image to obtain a frequency domain feature image; and to perform feature filtering on the frequency domain feature image to obtain a target feature image.

[0011] The present invention also provides a category estimation device, comprising: an initial image acquisition module for acquiring an initial image; the initial image including a target object to be processed; a spatial feature image acquisition module for extracting features from the initial image to obtain a spatial feature image; a frequency domain feature image acquisition module for processing the spatial feature image to obtain a frequency domain feature image; a target feature image acquisition module for filtering features from the frequency domain feature image to obtain a target feature image; an intermediate feature image acquisition module for performing dimensionality reduction processing on the target feature image to obtain an intermediate feature image; and a category estimation acquisition module for obtaining a category estimate of the target object based on the intermediate feature image.

[0012] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any of the above-described category estimation methods.

[0013] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the above-described category estimation methods.

[0014] The category estimation method, apparatus, electronic device, and storage medium provided by this invention involve: acquiring an initial image; the initial image including a target object to be processed; extracting features from the initial image to obtain a spatial feature image; processing the spatial feature image to obtain a frequency domain feature image; filtering features from the frequency domain feature image to obtain a target feature image; performing dimensionality reduction on the target feature image to obtain an intermediate feature image; and obtaining a category estimate of the target object based on the intermediate feature image. By converting the feature image from spatial features to frequency domain features, and performing feature filtering and dimensionality reduction on the frequency domain feature image, the category estimate of the target object can be accurately obtained, improving the accuracy of category estimation. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0016] Figure 1 This is one of the flowcharts illustrating the category estimation method provided by the present invention;

[0017] Figure 2 This is the second flowchart illustrating the category estimation method provided by the present invention;

[0018] Figure 3 This is one of the structural schematic diagrams of the category estimation method provided by the present invention;

[0019] Figure 4 This is the second schematic diagram of the structure of the category estimation method provided by the present invention;

[0020] Figure 5 This is the third schematic diagram of the structure of the category estimation method provided by the present invention;

[0021] Figure 6 This is a schematic diagram of the category estimation device provided by the present invention;

[0022] Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0024] The following is combined Figures 1-2 The category estimation method of the present invention is described. Taking the application of this method to a server as an example, the method includes the following steps:

[0025] Step 102: Obtain an initial image; the initial image includes the target object to be processed.

[0026] In this context, the target object refers to the object to be processed that exists within the initial image. For example, if the initial image includes a person, and that person is the object for which category estimation is required, then the object in the initial image is the target object.

[0027] Specifically, the server can obtain the initial image directly or indirectly.

[0028] In one embodiment, the server is connected to an image acquisition device, which can acquire an initial image and transmit the acquired initial image to the connected server; accordingly, the server obtains the aforementioned initial image.

[0029] Step 104: Extract features from the initial image to obtain a spatial feature image.

[0030] Spatial features refer to features that have spatial relationships. These features can enhance the ability to describe and distinguish image content.

[0031] Specifically, after obtaining the initial image, the server uses a spatial feature extraction method to extract spatial features from the initial image to obtain a spatial feature image.

[0032] In one embodiment, after obtaining the initial image, the server can use feature extraction methods such as HOG (Histogram of Oriented Gradient) or SIFT (Scale-invariant feature transform) to extract features from the initial image to obtain a spatial feature image. Understandably, feature extraction methods such as SURF (Speeded UpRobust Features) or ORB (Oriented Fast and Rotated Brief) can also be used to extract features from the initial image to obtain a spatial feature image.

[0033] Step 106: Process the spatial feature image to obtain the frequency domain feature image.

[0034] Frequency domain features refer to the characteristics of an image in the frequency domain.

[0035] Specifically, the server can convert a spatial feature image into a frequency domain feature image using a conversion method between spatial and frequency domain features.

[0036] In one embodiment, the server can convert a spatial feature image into a frequency domain feature image using a discrete Fourier transform.

[0037] Step 108: Perform feature filtering on the frequency domain feature image to obtain the target feature image.

[0038] Specifically, after the server obtains the frequency domain feature image, it can use feature filtering methods to filter out features from the frequency domain feature image to obtain the target feature image.

[0039] In one embodiment, the server uses a stored frequency domain filtering function to filter out features from the frequency domain feature image, removing those frequency domain feature images that are greater than or equal to a frequency threshold, and obtaining frequency domain feature images that are less than the frequency threshold.

[0040] In one embodiment, the module executing the methods in steps 106 and 108 can be identified as a low-frequency module, and the low-frequency module can be inserted between the convolutional and pooling layers of the neural network in a plug-and-play structure, thereby improving the robustness of the neural network while reducing memory and computation costs.

[0041] Step 110: Dimensionality reduction is performed on the target feature image to obtain an intermediate feature image.

[0042] Dimensionality reduction refers to transforming a target feature image into a data set in a high-dimensional space by increasing its dimensionality. Dimensionality reduction transforms the recognition of target feature images into a problem of recognizing feature representation vectors, thereby reducing computational complexity and achieving better image recognition results.

[0043] Specifically, after the server obtains the target feature image, it can perform dimensionality reduction processing on the target feature image to obtain an intermediate feature image.

[0044] In one embodiment, the server can use linear or nonlinear dimensionality reduction methods to process the target feature image to obtain an intermediate feature image.

[0045] In one embodiment, the server can input the target feature image into the pooling layer of a neural network for dimensionality reduction to obtain an intermediate feature image.

[0046] Step 112: Based on the intermediate feature image, obtain the category estimate of the target object.

[0047] Category estimation refers to the estimation of the category to which a target object belongs. For example, if the target object is an animal, the estimate of correctly identifying the target object as belonging to a certain animal category is 90%. In other words, this category estimation can also be understood as the probability of correctly identifying the target object.

[0048] Specifically, after obtaining the intermediate feature image, the server classifies the intermediate feature image to obtain a category estimate of the target object.

[0049] In one embodiment, the server can input the intermediate feature image into the fully connected layer of the neural network, and after processing by the fully connected layer, obtain the category estimate of the target object.

[0050] The above-described category estimation method involves: acquiring an initial image containing the target object; extracting features from the initial image to obtain a spatial feature image; processing the spatial feature image to obtain a frequency domain feature image; filtering features from the frequency domain feature image to obtain a target feature image; reducing the dimensionality of the target feature image to obtain an intermediate feature image; and obtaining a category estimate of the target object based on the intermediate feature image. By converting the feature image from spatial features to frequency domain features, and performing feature filtering and dimensionality reduction on the frequency domain feature image, the category estimate of the target object can be accurately obtained, thus improving the accuracy of category estimation.

[0051] In one embodiment, the frequency domain feature image includes a first frequency domain feature and a second frequency domain feature; filtering the frequency domain feature image to obtain a target feature image includes: filtering the first frequency domain feature in the frequency domain feature image according to a frequency threshold to obtain the target feature image; wherein the target feature image includes the second frequency domain feature.

[0052] Here, the frequency threshold refers to a critical value for frequency. Features greater than or equal to this threshold are considered to be in the first frequency domain, while features below this threshold are considered to be in the second frequency domain. It can be understood that frequency represents a distribution within an image.

[0053] Specifically, the server can use a low-pass filter to remove the first frequency domain feature, i.e., the high-frequency feature, from the frequency domain feature image to obtain the target feature image. Understandably, this target feature image is a low-frequency feature image.

[0054] In this embodiment, by filtering out the first frequency domain feature in the frequency domain feature image according to the frequency threshold, the target feature image is obtained, which can enhance the overall strength of the image.

[0055] In one embodiment, such as Figure 2 As shown, filtering out the first frequency domain feature in the frequency domain feature image yields the target feature image, which includes:

[0056] Step 202: Obtain the frequency domain filtering function.

[0057] Specifically, frequency domain filtering functions can smooth images and filter out noise.

[0058] In one embodiment, the server utilizes a Gaussian low-pass filter with an m*m kernel and transforms the Gaussian low-pass filter from the frequency domain to the spatial domain. The function G() of the Gaussian low-pass filter can be expressed as the formula:

[0059]

[0060] Where x represents the horizontal coordinate of an image pixel in the spatial domain, and y represents the vertical coordinate of an image pixel in the spatial domain.

[0061] In one embodiment, the server can obtain a frequency domain filtering plugin or script that has the same function as the frequency domain filtering function by acquiring built-in plugins or built-in scripts.

[0062] Step 204: Transform the frequency domain filtering function from the frequency domain to the spatial domain to obtain the spatial domain filtering function.

[0063] Specifically, after the server obtains the frequency domain filtering function, it can process the frequency domain filtering function by converting from the frequency domain to the spatial domain.

[0064] In one embodiment, the server uses the inverse discrete Fourier transform to convert the frequency domain filtering function obtained above to the spatial domain, thus obtaining the spatial domain filtering function.

[0065] Step 206: Use the spatial domain filtering function to filter out the first frequency domain feature in the frequency domain feature image to obtain the target feature image.

[0066] Specifically, after obtaining the spatial domain filtering function, the server filters out the first frequency domain feature in the frequency domain feature image to obtain the target feature image.

[0067] In one embodiment, the high-frequency features in the frequency domain feature image are filtered out using the function of the aforementioned Gaussian low-pass filter to obtain a low-frequency feature image. It is understood that the full-frequency features include both low-frequency and high-frequency features. Low-frequency features describe the smooth changes in the image's structure, while high-frequency features describe rapidly changing fine details. Low-frequency features are more domain-invariant than images containing full-frequency features, while high-frequency features contain domain-related information and can easily affect the alignment of data distribution.

[0068] In this embodiment, by obtaining a frequency domain filtering function, converting the frequency domain filtering function from the frequency domain to the spatial domain, a spatial domain filtering function is obtained. Using the spatial domain filtering function, the first frequency domain feature in the frequency domain feature image is filtered out to obtain the target feature image, thus achieving the goal of accurately obtaining the target feature image.

[0069] In one embodiment, processing a spatial feature image to obtain a frequency domain feature image includes: performing a discrete Fourier transform on the spatial feature image to obtain a frequency domain feature image; and performing feature filtering on the frequency domain feature image to obtain a target feature image includes: performing feature filtering on the frequency domain feature image to obtain a filtered feature image; and performing an inverse discrete Fourier transform on the filtered feature image to obtain the target feature image.

[0070] Specifically, after obtaining the spatial feature image, the server performs a discrete Fourier transform on the spatial feature image to obtain a frequency domain feature image; and after filtering out features from the frequency domain feature image, it uses an inverse discrete Fourier transform to convert the filtered feature image into the target feature image.

[0071] In this embodiment, a frequency domain feature image is obtained by performing a discrete Fourier transform on the spatial feature image, and a filtered feature image is obtained by filtering the frequency domain feature image. The target feature image is obtained by performing an inverse discrete Fourier transform on the filtered feature image. This achieves the goal of accurately obtaining the target feature image in the spatial domain after filtering out specific frequency features.

[0072] In one embodiment, extracting features from an initial image to obtain a spatial feature image includes: extracting features from the initial image using a convolutional layer of a neural network to obtain a spatial feature image.

[0073] Specifically, the server can use a neural network to extract features from the initial image. The initial image is input into the input layer of the neural network, and after the input layer is matrixed, features are extracted from the initial image in the convolutional layer to obtain a spatial feature image.

[0074] In this embodiment, by using the convolutional layer of a neural network to extract features from the initial image, a spatial feature image is obtained, which can achieve the goal of accurately extracting spatial feature images.

[0075] In one embodiment, the convolutional layer includes a strided convolutional layer; using the convolutional layer of the neural network to extract features from the initial image to obtain a spatial feature image includes: replacing the strided convolutional layer with a frequency module to extract features from the initial image to obtain a spatial feature image; wherein the frequency module is used to perform the step of processing the spatial feature image to obtain a frequency domain feature image; and performing feature filtering on the frequency domain feature image to obtain a target feature image.

[0076] Among them, a strided convolutional layer refers to a convolutional layer with a stride greater than 1 and a downsampling function. For example, if the previous convolution step convolves a 3*3 region from (0,0) to (2,2), the next convolution step is a 3*3 region from (2,0) to (4,2) or from (0,2) to (2,4).

[0077] Specifically, when the convolutional layer of the neural network includes a strided convolutional layer, the strided convolutional layer can be replaced with the frequency module described above, which performs the methods in steps 106 and 108.

[0078] In one embodiment, after the server identifies that there are strided convolutional layers in each layer of the neural network, it replaces the strided convolutional layer with the frequency module mentioned above and performs the methods in steps 106 and 108 above, that is, processes the spatial feature image to obtain a frequency domain feature image; and performs feature filtering on the frequency domain feature image to obtain a target feature image.

[0079] In this embodiment, by replacing the strided convolutional layer with a frequency module, feature extraction is performed on the initial image to obtain a spatial feature image. The frequency module is used to process the spatial feature image to obtain a frequency domain feature image. Feature filtering is performed on the frequency domain feature image to obtain a target feature image. This allows the network to focus more on low-frequency information. The frequency module can perform low-pass filtering on each input feature map, and the parameters in the frequency module are fixed.

[0080] In one embodiment, taking the aforementioned frequency module as an example, the low-frequency module (LFM) is constructed from a digital Gaussian low-pass filter. It can enhance the model's generalization performance in domain adaptation tasks by utilizing the inherent low-frequency information of the feature map. It is easy to implement and does not introduce additional hyperparameters. Domain adaptation of the LFM can be performed using the following two methods. Specifically, such as... Figure 3 As shown, Method 1 involves inserting LFM (Low-Frequency Modeling) after the convolutional layers and before the global pooling layers in a neural network, such as ResNet or VGGNet, to extract low-frequency information contained in the feature maps. This method ensures that the feature maps processed by the linear classification layers contain low-frequency information. Figure 4As shown, Method 2 replaces the 3x3 strided convolutional layer with LFM. To further extract low-frequency information from feature maps at different stages, this invention replaces the strided convolutional layer with LFM in the encoder network. By increasing the LFM constraint in the block (convolutional block), the network will pay more attention to low-frequency information. The diagram shows a comparison structure of strided convolution and a block equipped with strided convolution. Unlike strided convolution, LFM performs a low-pass filtering operation on each input feature map, and its parameters are fixed. The method in this embodiment, by inserting LFM into the neural network or using LFM to replace strided convolution in the block, can address the domain adaptation problem from a frequency perspective. Furthermore, LFM can be used as a plug-and-play unit, which can be combined with existing semi-supervised learning and data augmentation (UDA) methods to achieve better generalization performance. It can be integrated into various CNN backbones for different tasks, image classification, object detection, and instance segmentation, and all integrations outperform existing technologies or baselines with lower memory and computational costs.

[0081] In one embodiment, in the aforementioned neural network with inserted LFM, the neural network is trained using a training image set, where each image in the training set carries a label. Training stops when the number of iterations is reached or the loss reaches a preset loss value, resulting in a trained neural network model. A test image set without labels is then input into the trained neural network model, and the classification prediction result of the target object in the image is output.

[0082] In one embodiment, a trained neural network model can be used to test the test set images multiple times to obtain the statistical value of the classification prediction result each time, and the statistical value can be used as the prediction result of the target object in the image.

[0083] In one embodiment, the LFM design described above can be configured with a kernel of size m. Specifically, the digital low-pass filter employs a Gaussian low-pass filter with a kernel of size m×m. This Gaussian low-pass filter is then transformed from the frequency domain to the spatial domain, and the function definition of the digital spatial Gaussian low-pass filter G() is as follows:

[0084]

[0085] Specifically, such as Figure 5As shown, the spatial feature map obtained through the neural network is converted into frequency features using a Discrete Fourier Transform (DFT). When this frequency feature is input into a Gaussian low-pass filter, the high-frequency information in the output is filtered out. Finally, the output feature map is obtained through an Inverse Discrete Fourier Transform (IDFT). It can be understood that the filtering out of high-frequency information refers to the suppression of high-frequency information when the frequency value exceeds the cutoff frequency.

[0086] The category estimation device provided by the present invention is described below. The category estimation device described below can be referred to in correspondence with the category estimation method described above.

[0087] In one embodiment, a category estimation apparatus is provided, comprising: an initial image acquisition module for acquiring an initial image; the initial image including a target object to be processed; a spatial feature image acquisition module for extracting features from the initial image to obtain a spatial feature image; a frequency domain feature image acquisition module for processing the spatial feature image to obtain a frequency domain feature image; a target feature image acquisition module for filtering features from the frequency domain feature image to obtain a target feature image; an intermediate feature image acquisition module for performing dimensionality reduction processing on the target feature image to obtain an intermediate feature image; and a category estimation acquisition module for obtaining a category estimate of the target object based on the intermediate feature image.

[0088] In one embodiment, the frequency domain feature image includes a first frequency domain feature and a second frequency domain feature; the target feature image obtaining module includes a target feature image obtaining unit, configured to filter out the first frequency domain feature in the frequency domain feature image according to a frequency threshold to obtain the target feature image; wherein the target feature image includes the second frequency domain feature.

[0089] In one embodiment, the target feature image obtaining unit is used to acquire a frequency domain filtering function; convert the frequency domain filtering function from the frequency domain to the spatial domain to obtain a spatial domain filtering function; and use the spatial domain filtering function to filter out the first frequency domain features in the frequency domain feature image to obtain the target feature image.

[0090] In one embodiment, the frequency domain feature image acquisition module is used to perform a discrete Fourier transform on the spatial feature image to obtain the frequency domain feature image; to perform feature filtering on the frequency domain feature image to obtain a filtered feature image; and to perform an inverse discrete Fourier transform on the filtered feature image to obtain the target feature image.

[0091] In one embodiment, a spatial feature image obtaining module is used to extract features from the initial image using convolutional layers of a neural network to obtain a spatial feature image.

[0092] In one embodiment, a spatial feature image acquisition module is used to replace the strided convolutional layer with a frequency module to extract features from the initial image to obtain a spatial feature image; wherein, the frequency module is used to perform a step to process the spatial feature image to obtain a frequency domain feature image; and to perform feature filtering on the frequency domain feature image to obtain a target feature image.

[0093] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7 As shown, the electronic device may include a processor 710, a communications interface 720, a memory 730, and a communication bus 740, wherein the processor 710, the communications interface 720, and the memory 730 communicate with each other via the communication bus 740. The processor 710 can call logical instructions in the memory 730 to execute a category estimation method, which includes: acquiring an initial image; the initial image including a target object to be processed; extracting features from the initial image to obtain a spatial feature image; processing the spatial feature image to obtain a frequency domain feature image; filtering features from the frequency domain feature image to obtain a target feature image; performing dimensionality reduction processing on the target feature image to obtain an intermediate feature image; and obtaining a category estimate of the target object based on the intermediate feature image.

[0094] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0095] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the category estimation method provided by the above methods. The method includes: acquiring an initial image; the initial image including a target object to be processed; extracting features from the initial image to obtain a spatial feature image; processing the spatial feature image to obtain a frequency domain feature image; filtering features from the frequency domain feature image to obtain a target feature image; performing dimensionality reduction processing on the target feature image to obtain an intermediate feature image; and obtaining a category estimate of the target object based on the intermediate feature image.

[0096] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the category estimation method provided by the methods described above. The method includes: acquiring an initial image; the initial image including a target object to be processed; extracting features from the initial image to obtain a spatial feature image; processing the spatial feature image to obtain a frequency domain feature image; filtering features from the frequency domain feature image to obtain a target feature image; performing dimensionality reduction processing on the target feature image to obtain an intermediate feature image; and obtaining a category estimate of the target object based on the intermediate feature image.

[0097] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0098] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An image processing method, characterized in that, include: Obtain an initial image; the initial image includes the target object to be processed; Feature extraction is performed on the initial image to obtain a spatial feature image; The step of extracting features from the initial image to obtain a spatial feature image includes: The initial image is used to extract features using the convolutional layers of a neural network to obtain a spatial feature image; The convolutional layer includes a strided convolutional layer; the process of using the convolutional layer of the neural network to extract features from the initial image to obtain a spatial feature image includes: The strided convolutional layer is replaced with a frequency module to extract features from the initial image, resulting in a spatial feature image. The frequency module replaces the strided convolution in the convolutional block with a plug-and-play structure. The frequency module performs steps to process the spatial feature image to obtain a frequency domain feature image. Feature filtering is then performed on the frequency domain feature image to obtain a target feature image. The frequency module is a low-frequency module. The target feature image is then subjected to dimensionality reduction processing to obtain an intermediate feature image; Based on the intermediate feature image, the category estimate of the target object is obtained.

2. The image processing method according to claim 1, characterized in that, The frequency domain feature image includes a first frequency domain feature and a second frequency domain feature; the step of filtering the frequency domain feature image to obtain the target feature image includes: Based on a frequency threshold, the first frequency domain features in the frequency domain feature image are filtered out to obtain the target feature image; wherein, the target feature image includes the second frequency domain features.

3. The image processing method according to claim 1, characterized in that, The process of processing the spatial feature image to obtain the frequency domain feature image includes: The frequency domain feature image is obtained by performing a discrete Fourier transform on the spatial feature image. The step of filtering out features from the frequency domain feature image to obtain the target feature image includes: The frequency domain feature image is subjected to feature filtering to obtain a filtered feature image; The target feature image is obtained by performing an inverse discrete Fourier transform on the filtered feature image.

4. An image processing apparatus, characterized in that, include: An initial image acquisition module is used to acquire an initial image; the initial image includes the target object to be processed. A spatial feature image acquisition module is used to extract features from the initial image to obtain a spatial feature image. The feature extraction from the initial image to obtain the spatial feature image includes: using a convolutional layer of a neural network to extract features from the initial image to obtain the spatial feature image; the convolutional layer includes a strided convolutional layer; the feature extraction from the initial image using the convolutional layer of the neural network to obtain the spatial feature image includes: replacing the strided convolutional layer with a frequency module to extract features from the initial image to obtain the spatial feature image; the frequency module replaces the strided convolutional layer in the convolutional block with a plug-and-play structure; wherein the frequency module performs steps to process the spatial feature image to obtain a frequency domain feature image; and performs feature filtering on the frequency domain feature image to obtain a target feature image; wherein the frequency module is a low-frequency module. The intermediate feature image acquisition module is used to reduce the dimensionality of the target feature image to obtain an intermediate feature image; The category estimation module is used to obtain a category estimate of the target object based on the intermediate feature image.

5. The image processing apparatus according to claim 4, characterized in that, The frequency domain feature image includes a first frequency domain feature and a second frequency domain feature; The target feature image acquisition module includes: The target feature image obtaining unit is used to filter out the first frequency domain feature in the frequency domain feature image according to a frequency threshold to obtain the target feature image; wherein the target feature image includes the second frequency domain feature.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the image processing method as described in any one of claims 1 to 3.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the image processing method as described in any one of claims 1 to 3.

Images