Radio foreground source modeling method based on deep learning and Gaussian fitting

Abstract

The present invention relates to image target detection. The present invention provides a radio foreground source modeling method based on deep learning and Gaussian fitting. A batch of simulation data is generated using astronomical software. The data includes the spatial position, frequency position and flux information of the target. By modifying the feature extraction backbone network, adding a channel attention mechanism, and training the neural network, the detected target spatial position is thresholded, a connected domain is established, and a local peak is found. Then, a two-dimensional Gaussian fitting is performed, and then a one-dimensional Gaussian fitting is performed on the frequency position of the target. Finally, the two-dimensional Gaussian model and the one-dimensional Gaussian model are combined to obtain a three-dimensional model. The present invention can perform end-to-end spatial modeling and frequency modeling on the target in the radio observation data, can perform real-time end-to-end foreground source processing and analysis on the radio observation data, and output reliable important astronomical parameters such as right ascension, declination, frequency and flux of the foreground source.

Description

Technical Field

[0001] The present invention relates to radio astronomy image source target detection and Gaussian fitting iterative modeling, and more specifically, to a radio foreground source modeling method based on deep learning and Gaussian fitting. Background Art

[0002] Radio telescopes can observe in multiple different frequency ranges, covering the entire radio band from low frequency to high frequency. The number of channels in the data cube generated by different numbers of radio telescope arrays is also different, and each channel in these generated radio data cubes represents observation information of different frequencies, and the information between different frequencies of foreground sources is interrelated. Traditional target detection methods perform target search on two-dimensional images. For images observed at different frequencies, different parameters need to be set and then detected. The disadvantage of this method is that it cannot run in parallel, resulting in low efficiency, and the correlation of frequency information is not considered during detection. Summary of the invention

[0003] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a radio foreground source modeling method based on deep learning and Gaussian fitting. The invention can further effectively detect the data acquired by the radio interferometer array and establish a complete model based on the deep learning neural network.

[0004] To achieve the above object, the present invention provides the following technical solution, comprising the following steps:

[0005] A radio foreground source modeling method based on deep learning and Gaussian fitting is provided. A backbone extraction network is set for radio astronomy data to extract the spatial information, frequency distribution information and flux information of the target to be detected. The least squares method is used to perform two-dimensional and one-dimensional Gaussian multiple iterations on the foreground source space, frequency distribution information and flux information obtained by the neural network, and a two-dimensional spatial model and a one-dimensional frequency model of the foreground source are fitted, and then combined to finally output a three-dimensional model. The method specifically includes the following steps:

[0006] Step 1. Construct a radio observation image annotation dataset: Based on the radio observation principle, observation equipment performance and observation mode, and referring to the measured data style, a radio observation image annotation dataset is generated using simulation software. The annotation dataset includes the simulated observation image, the spatial position of the target in the image, and the frequency distribution position;

[0007] Step 2. According to the continuity of foreground sources in the radio observation image in the frequency direction, a deep convolutional network is established to obtain the spatial distribution position and frequency distribution position of the foreground source, where the spatial distribution position is marked with a two-dimensional detection frame and the frequency distribution position is marked with an effective energy distribution range;

[0008] Step 3. In the spatial dimension, perform local two-dimensional Gaussian fitting modeling on the target detected by the deep convolutional neural network to determine the position of the foreground source in the spatial dimension, obtain a two-dimensional Gaussian model of the foreground source, and convert the pixel coordinates of the foreground source, i.e., the horizontal coordinate and the vertical coordinate, into the right ascension coordinate and declination coordinate in the celestial coordinate system through celestial coordinate system conversion;

[0009] Step 4. Based on the spatial dimension detection in step 3, the energy integration result of the foreground source in each frequency band detection frame is calculated, and then a one-dimensional modeling is performed along the frequency direction using a multidimensional Gaussian function to obtain the distribution information of the foreground source in frequency, and then a one-dimensional Gaussian frequency model of the foreground source is constructed;

[0010] Step 5. Based on the results obtained in steps 2, 3 and 4, the Gaussian modeling results in the frequency direction and the two-dimensional Gaussian modeling results in the spatial direction are combined to obtain a three-dimensional model of the target in the radio data.

[0011] Furthermore, in step 1, in order to obtain the characteristic information of the foreground source at each frequency and the correlation at different frequencies, the structure in the entire feature extraction backbone network is modified so that the model trained by the convolutional neural network can learn the frequency association information of the source on the basis of learning the location information of the source, so that the neural network can extract both the characteristic information of the foreground source at a single frequency and the correlation characteristics of the foreground source at different frequencies.

[0012] Furthermore, in step 1, the structure of the entire feature extraction backbone network is modified as follows: first, the number of input channels of the first layer of convolution kernel is modified to automatically match the frequency of the input data, and then the number of output channels is set according to the original input-output ratio coefficient. At the same time, in order to avoid the phenomenon of network non-convergence during training, the threshold of the number of convolution kernel input channels is set.

[0013] Furthermore, in step 2, Gaussian fitting is performed for multiple iterations using the target spatial position information, peak brightness information, rotation angle, and observation beam information until the residual ratio between the fitted image and the original image is 1%, and the iteration is stopped to obtain a two-dimensional Gaussian model of all targets.

[0014] Furthermore, in step 3, a local two-dimensional Gaussian fitting is performed in the spatial dimension, and the target spatial position information obtained by the neural network is sliced ​​and extracted along the frequency direction to obtain the target position corresponding to each frequency, and then all pixels in each frequency slice are summed to obtain the frequency distribution characteristics of the target. The frequency distribution information is used to perform one-dimensional Gaussian fitting for multiple iterations until the residual ratio between the fitted curve and the original curve is 1%, and the iteration is stopped, and finally a one-dimensional Gaussian model of all targets is obtained.

[0015] In summary, the invention has the following beneficial effects:

[0016] Based on the traditional target detection neural network, the present invention further designs a backbone extraction network for radio astronomy data cube, which can not only extract the spatial information of the target, but also extract the frequency distribution information and flux information of the target.

[0017] The present invention uses the least squares method to perform two-dimensional and one-dimensional Gaussian multiple iteration fitting on the foreground source space, frequency and flux information extracted by the neural network to obtain the final two-dimensional space model and one-dimensional frequency model of the foreground source, and then combine them to finally output a three-dimensional model.

[0018] The present invention can perform real-time end-to-end foreground source processing and analysis on radio observation data, and output reliable foreground source positions, i.e. right ascension and declination, and important astronomical parameters of frequency and flux. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0020] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0021] It should be noted that, for the sake of convenience, the directions described below are consistent with the directions of the drawings themselves, but do not limit the structure of the present invention.

[0022] like Figure 1 As shown, the present invention discloses a method for modeling foreground targets in radio observation images based on a combination of deep learning and Gaussian fitting, and the implementation includes the following steps:

[0023] Step 1. According to the radio observation principle, observation mode and the performance of the radio array, the simulation software Oskar is used to first read the GLEAM (GaLactic and Extragalactic All-sky MWA) star catalog, then set the RA (right ascension), DEC (declination), flux and Ref_freq (reference frequency) variables to create a sky model, set the start frequency and frequency increment of the observation, generate a batch of radio observation images containing the target spatial position and frequency position, and form a labeled data set. The radio observation images are extracted from the radio observation images in the labeled data set to extract the simulated observation images, the spatial position and frequency distribution position of the target in the image. Due to the particularity of the frequency information of the radio observation data, the foreground source can only be observed in a specific frequency band, and the frequency bands that can be observed by different foreground sources are also different. Therefore, a deep convolutional network is established for the characteristics of the foreground source. The input labels of the deep convolutional neural network add frequency labels on the basis of the original horizontal and vertical coordinates, so that the network can directly predict the spatial position and frequency position of the output target; in order to better obtain the characteristic information of the data at each frequency and the correlation of different frequencies, the feature extraction backbone network is modified as follows: first, the number of input channels of the first layer of convolution kernel is modified to automatically match the frequency number of input data, that is, the number of channels, and then the number of output channels is set according to the original input-output ratio coefficient. In order to prevent too many parameters in the network from causing the network to not converge during training, the threshold of the number of convolution kernel input channels is set. A channel attention mechanism is added on the basis of the original feature extraction backbone network to extract the frequency distribution information of the data and the frequency-related information of each foreground source. The specific modifications are as follows: The dimension of the input feature map is C×H×W, where C represents the number of channels, H and W represent the height and width respectively. For the input feature map, firstly, the global average pooling operation is performed to average pool in the spatial dimension to obtain a global average pooling feature vector with a dimension of C. The global average pooling feature vector is input into a fully connected layer to generate an intermediate feature vector with a dimension of C. The intermediate feature vector is input into two fully connected layers respectively, so that the deep convolutional network can learn the complex nonlinear relationship between position and frequency in the data cube, and then two weight vectors with a dimension of C are obtained, which are used to represent the importance weight and suppression weight of the frequency respectively. The frequency importance weight vector is multiplied with the input feature map to obtain the weighted feature map, and the weighted feature map is input into a sigmoid function to map the frequency suppression weight vector to a probability value in the range of [0,1]. The frequency suppression weight vector is multiplied with the weighted feature map to obtain the final channel attention feature map, which is used to extract features. The final output of the neural network includes the spatial position information, brightness information and frequency information of the target.

[0024] Step 2. According to the continuity of the foreground source in the frequency direction in the radio observation image, a deep convolutional network is established to obtain the spatial distribution position and frequency distribution position of the foreground source, where the spatial distribution position is marked by a two-dimensional detection frame, and the frequency distribution position is marked by an effective energy distribution range. The image is sliced ​​based on the target spatial position information, peak brightness information, rotation angle and observation beam information obtained in step 1, and the local target image is extracted. The pixel values ​​in the local target image that are less than the standard deviation σ are set to 0, and the pixel values ​​that are greater than the standard deviation (σ) are retained. The non-zero elements are connected into regions, and the local image is processed to extract the connected regions. Then, the maximum value filtering method is used to search for peaks, and the results are iterated by two-dimensional Gaussian fitting until the residual ratio between the fitted image and the original image is 1%, and the two-dimensional spatial model of the target is obtained.

[0025] Step 3. In the spatial dimension, perform local two-dimensional Gaussian fitting modeling on the target detected by the deep convolutional neural network to determine the position of the foreground source in the spatial dimension, and obtain a two-dimensional Gaussian model of the foreground source. Through the celestial coordinate system conversion, the pixel coordinates of the foreground source, i.e. the horizontal coordinate and the vertical coordinate, are converted into the right ascension coordinate and declination coordinate in the celestial coordinate system. The target spatial position information obtained in step 1 is sliced ​​and extracted along the frequency direction for each frequency corresponding to the target position, and then all pixels in each frequency slice are summed to obtain the frequency distribution characteristics of the target, and the frequency information is denoised to avoid the influence of noise on the subsequent Gaussian fitting, and then the one-dimensional Gaussian fitting method is used for fitting iteration to finally obtain the one-dimensional frequency distribution model of the target.

[0026] Step 4. Based on the spatial dimension detection in step 3, the energy integration result of the foreground source in each frequency band bounding box is calculated, and then a one-dimensional modeling is performed along the frequency direction using a multidimensional Gaussian function to obtain the distribution information of the foreground source in frequency, and then a one-dimensional Gaussian frequency model of the foreground source is constructed;

[0027] Step 5. Combine the two-dimensional Gaussian modeling results in the spatial direction and the one-dimensional Gaussian modeling results in the frequency direction to finally obtain a three-dimensional model of the foreground source in the radio data.

[0028] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A radio foreground source modeling method based on deep learning and Gaussian fitting, characterized in that: A backbone extraction network is set up for radio astronomy data to extract the spatial information, frequency distribution information and flux information of the target to be detected. The least squares method is used to perform two-dimensional and one-dimensional Gaussian multiple iterations on the foreground source space, frequency distribution information and flux information obtained by the neural network, and the two-dimensional spatial model and one-dimensional frequency model of the foreground source are fitted and combined to finally output a three-dimensional model. The specific steps include: Step 1. Construct a radio observation image annotation dataset: Based on the radio observation principle, observation equipment performance and observation mode, and referring to the measured data style, a radio observation image annotation dataset is generated using simulation software. The annotation dataset includes the simulated observation image, the spatial position of the target in the image, and the frequency distribution position; Step 2. According to the continuity of foreground sources in the radio observation image in the frequency direction, a deep convolutional network is established to obtain the spatial distribution position and frequency distribution position of the foreground source, where the spatial distribution position is marked with a two-dimensional detection frame and the frequency distribution position is marked with an effective energy distribution range; Step 3. In the spatial dimension, perform local two-dimensional Gaussian fitting modeling on the target detected by the deep convolutional neural network to determine the position of the foreground source in the spatial dimension, obtain a two-dimensional Gaussian model of the foreground source, and convert the pixel coordinates of the foreground source, i.e., the horizontal coordinate and the vertical coordinate, into the right ascension coordinate and declination coordinate in the celestial coordinate system through celestial coordinate system conversion; Step 4. Based on the spatial dimension detection in step 3, the energy integration result of the foreground source in each frequency band detection frame is calculated, and then a one-dimensional modeling is performed along the frequency direction using a multidimensional Gaussian function to obtain the distribution information of the foreground source in frequency, and then a one-dimensional Gaussian frequency model of the foreground source is constructed; Step 5. Based on the results obtained in steps 2, 3 and 4, the Gaussian modeling results in the frequency direction and the two-dimensional Gaussian modeling results in the spatial direction are combined to obtain a three-dimensional model of the target in the radio data.

2. The radio foreground source modeling method based on deep learning and Gaussian fitting according to claim 1, characterized in that: In the step 1, in order to obtain the characteristic information of the foreground source at each frequency and the correlation at different frequencies, the structure in the entire feature extraction backbone network is modified so that the model trained by the convolutional neural network can learn the frequency association information of the source on the basis of learning the location information of the source, so that the neural network can extract both the characteristic information of the foreground source at a single frequency and the correlation characteristics of the foreground source at different frequencies.

3. The radio foreground source modeling method based on deep learning and Gaussian fitting according to claim 2, characterized in that: In step 1, the structure of the entire feature extraction backbone network is modified as follows: first, the number of input channels of the first layer of convolution kernel is modified to automatically match the frequency number of the input data, and then the number of output channels is set according to the original input-output ratio coefficient. At the same time, in order to avoid the phenomenon of network non-convergence during training, the threshold of the number of convolution kernel input channels is set.

4. The radio foreground source modeling method based on deep learning and Gaussian fitting according to claim 1, characterized in that: In step 2, Gaussian fitting is performed multiple times by using the target spatial position information, peak brightness information, rotation angle, and observation beam information until the residual ratio between the fitted image and the original image is 1%, and the iteration is stopped to obtain a two-dimensional Gaussian model of all targets.

5. The radio foreground source modeling method based on deep learning and Gaussian fitting according to claim 1, characterized in that: In step 3, a local two-dimensional Gaussian fitting is performed in the spatial dimension, and the target spatial position information obtained by the neural network is sliced ​​and extracted along the frequency direction to obtain the target position corresponding to each frequency, and then all pixels in each frequency slice are summed to obtain the frequency distribution characteristics of the target. The frequency distribution information is used to perform one-dimensional Gaussian fitting for multiple iterations until the residual ratio between the fitted curve and the original curve is 1%, and the iteration is stopped, and finally a one-dimensional Gaussian model of all targets is obtained.

Images