Pulsar candidate body generation method and system based on improved DCGAN

Abstract

According to the pulsar candidate body generation method and system based on the improved DCGAN, pulsar candidate body images are generated through the improved DCGAN, channel information of different scales is extracted, and then channel attention weights are generated through convolution processing of different scales; multi-scale features of output features are extracted, and then a space attention map is obtained based on different pooling operation processing; local details and global features of the image can be enhanced, so that the overall quality and detail performance of the generated image are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of image processing technology, and in particular to a pulsar candidate generation method and system based on an improved DCGAN. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Pulsars are a type of highly compact celestial body known for their rapid rotation and regular radio pulse signals. Studying pulsars is of great significance for understanding the basic physical phenomena in the universe. Pulsars can be used as precise astrometric tools to study phenomena such as gravity, interstellar medium, magnetic fields and interstellar dust, and to explore the evolutionary history of the universe, the structure of the interstellar medium and the properties of the interstellar magnetic field.

[0004] With the completion of the Five-hundred-meter Aperture Spherical Radio Telescope (FAST), the amount of pulsar observation data has exploded. However, there is a class imbalance problem in the pulsar candidate dataset. This imbalance will cause the classification model to tend to predict the larger number of categories during training, while ignoring the smaller number of categories, thus affecting the model's ability to recognize the minority categories.

[0005] In the field of pulsar candidate classification, the positive-negative class imbalance problem is an important challenge, which refers to the significant difference in the number of samples of different categories in the dataset. This imbalance will cause the classification model to tend to predict the category with a larger number during training, while ignoring the category with a smaller number, thus affecting the model's ability to recognize the minority category.

[0006] Therefore, how to generate more accurate pulsar candidate images and solve the imbalance in the number of samples of different categories in the classification of pulsar candidates is a problem that needs to be solved at present. Summary of the invention

[0007] In order to overcome the above-mentioned deficiencies of the prior art, the present invention provides a pulsar candidate generation method and system based on improved DCGAN, which significantly improves the overall quality and detail performance of the pulsar candidate generated image by enhancing the local details and global features of the image.

[0008] In order to achieve the above object, the present invention adopts the following technical solution:

[0009] In a first aspect, the present invention provides a pulsar candidate generation method based on an improved DCGAN, comprising:

[0010] Acquire pulsar candidate images and perform preprocessing;

[0011] The preprocessed pulsar candidate image is input into the trained improved DCGAN model to obtain the pulsar candidate generated image;

[0012] The operation of the generator of the improved DCGAN model on the input image is specifically as follows:

[0013] Perform deconvolution operation on the input image to obtain the input feature map;

[0014] Extracting channel information of different scales from the input feature map to obtain different feature map representations; fusing the different feature map representations and subjecting them to convolution processing of different scales to generate channel attention weights, and fusing them with the input feature map to obtain an output feature map;

[0015] Performing convolution operations of different scales on the output feature map to extract multi-scale features; performing maximum pooling and average pooling on the extracted multi-scale features and fusing them to obtain a spatial attention map;

[0016] The spatial attention map is subjected to deconvolution processing to obtain a pulsar candidate generation image.

[0017] In a second aspect, the present invention provides a pulsar candidate generation method and system based on an improved DCGAN, including:

[0018] An acquisition module is configured to: acquire pulsar candidate images and perform preprocessing;

[0019] A generation module is configured to: input the preprocessed pulsar candidate image into the trained improved DCGAN model to obtain a pulsar candidate generated image;

[0020] The operation of the generator of the improved DCGAN model on the input image is specifically as follows:

[0021] Perform deconvolution operation on the input image to obtain the input feature map;

[0022] Extracting channel information of different scales from the input feature map to obtain different feature map representations; fusing the different feature map representations and subjecting them to convolution processing of different scales to generate channel attention weights, and fusing them with the input feature map to obtain an output feature map;

[0023] Performing convolution operations of different scales on the output feature map to extract multi-scale features; performing maximum pooling and average pooling on the extracted multi-scale features and fusing them to obtain a spatial attention map;

[0024] The spatial attention map is subjected to deconvolution processing to obtain a pulsar candidate generation image.

[0025] In a third aspect, the present invention provides an electronic device comprising a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method described in the first aspect is performed.

[0026] In a fourth aspect, the present invention provides a computer-readable storage medium for storing computer instructions, wherein when the computer instructions are executed by a processor, the method described in the first aspect is performed.

[0027] In a fifth aspect, the present invention provides a computer program product, comprising a computer program, which, when executed by a processor, implements the method described in the first aspect.

[0028] One or more of the above technical solutions have the following beneficial effects:

[0029] In the present invention, an improved DCGAN model is used to generate pulsar candidate images. Channel information of different scales is extracted, and then convolution processing of different scales is used to generate channel attention weights. Multi-scale features of output features are extracted, and then processed based on different pooling operations to obtain a spatial attention map. The local details and global features of the image can be enhanced, thereby significantly improving the overall quality and detail performance of the generated image.

[0030] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0032] Figure 1 This is an experimental flow chart of Example 1 of the present invention.

[0033] Figure 2 This is a structural diagram of the EDPSNet model in Example 1 of the present invention.

[0034] Figure 3 This is a structural diagram of the CSA module designed in Example 1 of the present invention. DETAILED DESCRIPTION

[0035] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0036] It should be noted that the terms used herein are for describing specific embodiments only and are not intended to be limiting of exemplary embodiments according to the present invention.

[0037] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0038] Embodiment 1

[0039] This embodiment discloses a method for generating pulsar candidates based on an improved DCGAN, including:

[0040] Acquire pulsar candidate images and perform preprocessing;

[0041] The preprocessed pulsar candidate image is input into the trained improved DCGAN model to obtain the pulsar candidate generated image;

[0042] The operation of the generator of the improved DCGAN model on the input image is specifically as follows:

[0043] Perform deconvolution operation on the input image to obtain the input feature map;

[0044] Extracting channel information of different scales from the input feature map to obtain different feature map representations; fusing the different feature map representations and subjecting them to convolution processing of different scales to generate channel attention weights, and fusing them with the input feature map to obtain an output feature map;

[0045] Performing convolution operations of different scales on the output feature map to extract multi-scale features; performing maximum pooling and average pooling on the extracted multi-scale features and fusing them to obtain a spatial attention map;

[0046] The spatial attention map is subjected to deconvolution processing to obtain a pulsar candidate generation image.

[0047] This embodiment proposes an improved DCGAN network, namely, the EDPSNet network. The EDPSNet network adopts DCGAN combined with a multi-scale Channel-Spatial Attention (CSA) module and Spectral normalization (SN) technology.

[0048] DCGAN uses convolutional layers and batch normalization layers to improve the quality of image generation. CSA is an attention module designed in this embodiment, which enhances the local details and global features of the image through multi-scale feature extraction, thereby significantly improving the overall quality and detail performance of the generated image. SN is a regularization technique that stabilizes the training process and prevents mode collapse by constraining the weight matrix in the network. This embodiment is committed to generating images through an improved DCGAN to solve the class imbalance problem in the pulsar candidate dataset, thereby significantly improving the classification accuracy and efficiency of pulsar candidates.

[0049] Combine the following Figure 1-Figure 2 The contents involved in the pulsar candidate generation method based on the improved DCGAN proposed in this embodiment are described in detail.

[0050] Step 1: Dataset preprocessing.

[0051] Step 11: Use the pulsar candidate dataset provided by FAST.

[0052] Before conducting the pulsar candidate generation experiment, the data set was first preprocessed. In the long time series data, when the pulsar signal was identified as a periodic signal, the pre-folding technique was used to generate diagnostic graphs. However, not all periodic signals originate from pulsars, so further analysis is needed to verify these candidates. The candidate pulsars are folded to produce a series of diagnostic sub-graphs, which are stored in the form of pfd format files. These files contain features such as pulse profiles, time phase diagrams, frequency phase diagrams, and DM curves.

[0053] Step 12: When determining the model features, due to the high instability of the dispersion correlation curve, this embodiment decides not to use the pulse profile and DM curve features, but chooses to use only the time phase diagram and the frequency phase diagram for subsequent analysis. This choice helps to focus on the most informative features, thereby improving the efficiency and accuracy of the model.

[0054] Step 13: After determining the features to be used, the data was further preprocessed. Since the time phase map and the frequency phase map have significant differences in size and scale, if not processed, it may have a negative impact on the learning effect and performance of the model. To address this problem, the time phase map and the frequency phase map were uniformly processed in size and scale before model training, and all images were resized to 64×64 pixels. This processing step not only eliminates unfavorable factors, but also ensures the consistency of data input, thereby improving the efficiency of model training and the quality of image generation.

[0055] Step 2: Build the EDPSNet model.

[0056] like Figure 3 As shown, step 21: the CSA module mainly includes two parts: a multi-scale channel attention module CA and a multi-scale spatial attention module SA.

[0057] The CSA module takes an intermediate vector X = [x1, x2, ..., x C ]∈R C×H×W As input, C is the number of channels, H is the height, and W is the width, and generates an enhanced representation Y = [y1, y2, ..., y C ].

[0058] CA introduces multi-scale pooling and convolution operations. First, the input feature map X is subjected to the maximum pooling layer and the average pooling layer to extract channel information of different scales, and two sets of different feature map representations are obtained, namely the first feature map and the second feature map. Then, the first feature map and the second feature map are convolved through the 3×3 convolution layer and the 5×5 convolution layer, respectively, and processed by the batch normalization layer after convolution. Finally, the channel attention weights are generated through the 3×3 convolution layer and the Sigmoid activation function, and the 5×5 convolution layer and the Sigmoid activation function, respectively, and weighted fused with the input feature map X to obtain C(X), the formula of which is as follows:

[0059]

[0060] Among them, C(X) represents the final output feature map, X represents the input feature map, and W 3×3 and W 5×5 They represent the channel attention weights obtained by 3×3 convolution and Sigmoid activation function, 5×5 convolution and Sigmoid activation function, respectively. Represents element-wise multiplication.

[0061] SA introduces multi-scale convolution and pooling operations. First, the feature map Y output by CA is input, and multi-scale feature extraction is performed through 3×3 convolution layer and 5×5 convolution layer respectively. Then, the feature map after 3×3 convolution is subjected to maximum pooling operation, and the feature map after 5×5 convolution is subjected to average pooling operation. After the maximum pooling operation and the average pooling operation, batch normalization is performed to improve the convergence speed and stability of the model; these feature maps are then fused by element-by-element multiplication, and finally processed through 1×1 convolution layer and Sigmoid activation function to obtain the spatial attention map S(Y), whose formula is as follows:

[0062]

[0064] Among them, F fusion represents the fused feature map, They represent the feature maps after 3×3 convolution and maximum pooling, and the feature maps after 5×5 convolution and average pooling, Y represents the feature map output by CA, and S(Y) represents the final output feature map. Finally, the output feature map CS containing CA and SA is obtained:

[0065]

[0066] Step 22: SN controls the Lipschitz constant of the network layer by adjusting the weight matrix to avoid the gradient explosion or vanishing problem. SN is based on the spectral norm of the weight matrix A, which is defined as follows:

[0067]

[0068] Among them, ∥·∥2 represents the L2 norm, and h is an arbitrary non-zero vector. By dividing the weight matrix A by its spectral norm σ(A), SN ensures that the Lipschitz constant after layer mapping is controlled, that is, the normalized weight matrix satisfy:

[0069]

[0070] This control can mitigate the aggressive behavior of the discriminator during training, making gradient updates smoother and avoiding gradient instability during activation and back-propagation of network layers. Through this normalization operation, the explosion or disappearance of gradients during training can be suppressed, thereby stabilizing the parameter update process. SN also ensures that features maintain a consistent scale during propagation of each layer, which is crucial to improving training stability and generation quality.

[0071] Step 3: Train the EDPSNet model

[0072] The weights of the generator and discriminator should be initialized with small random values ​​to avoid large gradients in the early stages of training, which can lead to unstable training. During training, the generator and discriminator are not updated simultaneously. First, the generator is fixed and the discriminator is optimized so that it learns to distinguish between real images and generated images. Then, the discriminator is fixed and the generator is optimized so that the images generated by the generator are increasingly difficult for the discriminator to recognize.

[0073] In order to effectively train the improved DCGAN for image generation tasks, the Wasserstein function is used as the loss function of the improved DCGAN network. The Wasserstein distance not only alleviates the gradient vanishing problem, but also reduces the risk of gradient explosion through more stable gradient calculation, making the DCGAN training process smoother and more stable.

[0074] SN is a regularization technique that stabilizes the training process and prevents mode collapse by constraining the weight matrix in the network.

[0075] The advantage of using Wasserstein function and SN together is that Wasserstein function provides a smoother and more meaningful loss function landscape, while SN ensures the stability and convergence of the discriminator.

[0076] Step 31: In the generator, the random noise z passes through the fully connected layer and is reshaped into the initial feature map The initial feature map goes through a series of deconvolution layers, batch normalization, and Relu activation functions to gradually enlarge the feature map size and generate detailed features. The deconvolution operation generates feature maps in turn. In the feature generation process, the feature map F2 is processed by the CSA module to obtain F3 is then processed by the CSA module to obtain the feature map F4, which is then input into the next deconvolution layer for further processing, and finally passes through the last deconvolution layer and the Tanh activation function to output the generated image

[0077] By combining CSA, the DCGAN generator significantly improves the quality and detail of the generated images, making it perform better in image generation tasks. CSA helps the generator better retain and express image detail features while capturing global structural information, thereby generating higher quality and more realistic images.

[0078] Step 32: The discriminator part starts with the input real or generated image, and goes through a series of convolutional layers (Conv), batch normalization (BN), and LeakyReLU activation functions to gradually reduce the feature map size and extract high-dimensional features. After each layer of convolution operation, the weight matrix is ​​normalized by SN to ensure the stability of the training process. The convolution operation reduces the feature map size to Finally, a high-dimensional feature map with a size of 512×4×4 is obtained. The high-dimensional feature map is activated by Flatten and Sigmoid functions to output the identification result. By applying SN to the DCGAN discriminator, the stability of the training process can be significantly improved, and the phenomenon of gradient explosion or disappearance can be reduced, thereby improving the overall performance of the model and the quality of the generated images.

[0079] Step 4: Verify the experimental results

[0080] Step 41: Visual observation is an intuitive way to verify the quality of generated images. The performance of the model can be evaluated by manually observing the sharpness, realism, diversity and other characteristics of the generated images. If the generated images have realistic details, textures and structures, it means that the model is well trained. Some samples of generated images can be saved periodically during the training process and manually evaluated to understand the improvement of the model.

[0081] Step 42: During the verification process, a FID threshold can be set. When the FID of the generated image is less than the threshold, the weight file of the model is saved. This indicates that the generator has learned a good enough representation and the quality of the generated image has reached a certain standard.

[0082] Step 43: The quality of the generated image can be evaluated using a pre-trained image classifier. This is done by calculating the classification accuracy of the generated image and the real image on the classifier. The higher the classification accuracy of the generated image, the better the quality of the generated image.

[0083] Embodiment 2

[0084] The purpose of this embodiment is to provide a pulsar candidate generation system based on an improved DCGAN, including:

[0085] An acquisition module is configured to: acquire pulsar candidate images and perform preprocessing;

[0086] A generation module is configured to: input the preprocessed pulsar candidate image into the trained improved DCGAN model to obtain a pulsar candidate generated image;

[0087] The operation of the generator of the improved DCGAN model on the input image is specifically as follows:

[0088] Perform deconvolution operation on the input image to obtain the input feature map;

[0089] Extracting channel information of different scales from the input feature map to obtain different feature map representations; fusing the different feature map representations and subjecting them to convolution processing of different scales to generate channel attention weights, and fusing them with the input feature map to obtain an output feature map;

[0090] Performing convolution operations of different scales on the output feature map to extract multi-scale features; performing maximum pooling and average pooling on the extracted multi-scale features and fusing them to obtain a spatial attention map;

[0091] The spatial attention map is subjected to deconvolution processing to obtain a pulsar candidate generation image.

[0092] In further embodiments, there is also provided:

[0093] An electronic device includes a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method described in Embodiment 1 is performed. For the sake of brevity, no further description is given here.

[0094] It should be understood that in this embodiment, the processor may be a central processing unit CPU, and the processor may also be other general-purpose processors, digital signal processors DSP, application-specific integrated circuits ASIC, off-the-shelf programmable gate arrays FPGA or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0095] The memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0096] A computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by a processor, the method described in embodiment 1 is completed.

[0097] The method in the first embodiment can be directly embodied as a hardware processor, or a combination of hardware and software modules in the processor. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware. To avoid repetition, it will not be described in detail here.

[0098] A computer program product includes a computer program, and when the computer program is executed by a processor, the method described in the first embodiment is implemented.

[0099] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer executable instructions, such as instructions included in a program module, which are executed in a device on a real or virtual processor of the target to perform the process / method as described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc. that perform specific tasks or implement specific abstract data types. In various embodiments, the functions of program modules can be combined or divided between program modules as needed. Machine executable instructions for program modules can be executed in local or distributed devices. In distributed devices, program modules can be located in local and remote storage media.

[0100] The computer program code for implementing the method of the present invention can be written in one or more programming languages. These computer program codes can be provided to the processor of a general-purpose computer, a special-purpose computer or other programmable data processing device, so that the program code, when executed by the computer or other programmable data processing device, causes the function / operation specified in the flow chart and / or block diagram to be implemented. The program code can be executed completely on a computer, partially on a computer, as an independent software package, partially on a computer and partially on a remote computer or completely on a remote computer or server.

[0101] In the context of the present invention, computer program codes or related data may be carried by any appropriate carrier to enable a device, apparatus or processor to perform the various processes and operations described above. Examples of carriers include signals, computer readable media, etc. Examples of signals may include electrical, optical, radio, acoustic or other forms of propagation signals, such as carrier waves, infrared signals, etc.

[0102] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0103] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A pulsar candidate generation method based on improved DCGAN, characterized in that: include: Acquire pulsar candidate images and perform preprocessing; The preprocessed pulsar candidate image is input into the trained improved DCGAN model to obtain the pulsar candidate generated image; The operation of the generator of the improved DCGAN model on the input image is specifically as follows: Perform deconvolution operation on the input image to obtain the input feature map; Extracting channel information of different scales from the input feature map to obtain different feature map representations; fusing the different feature map representations and subjecting them to convolution processing of different scales to generate channel attention weights, and fusing them with the input feature map to obtain an output feature map; Performing convolution operations of different scales on the output feature map to extract multi-scale features; performing maximum pooling and average pooling on the extracted multi-scale features and fusing them to obtain a spatial attention map; The spatial attention map is subjected to deconvolution processing to obtain a pulsar candidate generation image.

2. The pulsar candidate generation method based on improved DCGAN according to claim 1, characterized in that: The time phase map and the frequency phase map of the pulsar candidate are selected and preprocessed as the input images of the trained improved DCGAN model. The preprocessing is specifically: unifying the size and dimension of the time phase map and the frequency phase map.

3. The pulsar candidate generation method based on improved DCGAN according to claim 1, characterized in that: The input feature map is processed using a multi-scale channel attention module, specifically: The input feature map is processed by maximum pooling and average pooling respectively to extract channel information of different scales and obtain different feature map representations; The different feature maps are represented by convolution layers of different scales for convolution operation, and the results of the convolution operation are processed by batch normalization layer; The results of batch normalization are processed through convolutional layers of different scales and corresponding activation functions to generate channel attention weights; The channel attention weight and the input feature map are weighted fused to obtain an output feature map.

4. The pulsar candidate generation method based on improved DCGAN according to claim 1, characterized in that: The output feature map is processed using a multi-scale spatial attention module, specifically: Use convolutions of different scales to extract multi-scale features from the output feature map; The extracted features of different scales are subjected to maximum pooling and average pooling respectively, and the pooled results are batch normalized respectively; The results of batch normalization are multiplied element by element, and then processed by convolution and activation function in sequence to obtain the spatial attention map.

5. The pulsar candidate generation method based on improved DCGAN according to claim 1, characterized in that: In the improved DCGAN model training, the Wasserstein function is used as the loss function.

6. The method for generating pulsar candidates based on improved DCGAN according to claim 1, characterized in that: Based on SN technology, the Lipschitz constant of the network layer is controlled by adjusting the weight matrix to ensure that the Lipschitz constant after layer mapping is controlled so that the normalized weight matrix satisfies: in, is the normalized weight matrix, A is the original weight matrix, and σ(A) is the spectral norm of the original weight matrix.

7. The pulsar candidate generation system based on improved DCGAN is characterized by: include: An acquisition module is configured to: acquire pulsar candidate images and perform preprocessing; A generation module is configured to: input the preprocessed pulsar candidate image into the trained improved DCGAN model to obtain a pulsar candidate generated image; The operation of the generator of the improved DCGAN model on the input image is specifically as follows: Perform deconvolution operation on the input image to obtain the input feature map; Extracting channel information of different scales from the input feature map to obtain different feature map representations; fusing the different feature map representations and subjecting them to convolution processing of different scales to generate channel attention weights, and fusing them with the input feature map to obtain an output feature map; Performing convolution operations of different scales on the output feature map to extract multi-scale features; performing maximum pooling and average pooling on the extracted multi-scale features and fusing them to obtain a spatial attention map; The spatial attention map is subjected to deconvolution processing to obtain a pulsar candidate generation image.

8. An electronic device, characterized in that: The method comprises a memory and a processor and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method according to any one of claims 1 to 6 is completed.

9. A computer-readable storage medium, characterized in that: Used to store computer instructions, which, when executed by a processor, complete the method described in any one of claims 1 to 6.

10. A computer program product, characterized in that The invention comprises a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 6.

Images