CN-122017828-A - Single-bit millimeter wave imaging method based on range migration enhanced depth expansion network

CN122017828ACN 122017828 ACN122017828 ACN 122017828ACN-122017828-A

Abstract

A single-bit millimeter wave imaging method based on a range migration enhanced depth expansion network comprises the steps of generating a training data set containing simulated single-bit echo data and corresponding true value images, and constructing a data set containing simulated single-bit echo data and corresponding true value images The method comprises the steps of a depth neural network of cascade stages, a depth expansion network, a target image and a training data set, wherein an iterative imaging algorithm based on a range migration operator is expanded into a network hierarchical structure, each stage of the depth expansion network comprises a hyperbolic tangent physical gradient module and a space perception self-adaptive non-convex regularization module which are sequentially connected, the physical consistency of an input image is updated by the hyperbolic tangent physical gradient module in each stage of the depth expansion network to generate the intermediate image, the intermediate image is input into the space perception self-adaptive non-convex regularization module to be subjected to denoising processing, the training data set is used for training the depth expansion network, actual measurement single-bit millimeter wave echo data to be processed is input into the trained depth expansion network, and the target image is output.

Inventors

YANG AFENG
DAI KAIWEI
GE SHAODI

Assignees

杭州电子科技大学

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (9)

1. The single-bit millimeter wave imaging method based on the range migration enhanced depth expansion network is characterized by comprising the following steps of: S1, generating a training data set containing simulated single-bit echo data and corresponding truth images; S2, constructing and comprising The depth neural network of each cascade stage expands an iterative imaging algorithm based on a range migration operator into a network hierarchical structure, wherein each stage in the depth expansion network comprises a hyperbolic tangent physical gradient module and a space perception self-adaptive non-convex regularization module which are sequentially connected; S3, at each stage of the deep expansion network, using a hyperbolic tangent physical gradient module to update the physical consistency of the input image, and generating an intermediate image; s4, inputting the intermediate image into a space perception self-adaptive non-convex regularization module for denoising; S5, training the deep expansion network by using the training data set, inputting the actually measured single-bit millimeter wave echo data to be processed into the trained deep expansion network, and outputting a target image.
2. The method for single-bit millimeter wave imaging based on the range migration enhanced depth expansion network according to claim 1, wherein S1 comprises injecting complex gaussian white noise with a dynamic signal-to-noise ratio into an original echo as a jitter source, so that single-bit symbol data statistically responds monotonically to amplitude values, and analog single-bit echo data is generated.
3. The single-bit millimeter wave imaging method based on the range migration enhanced deep expansion network according to claim 1, wherein S2 comprises embedding fixed forward RM mapping and reverse RM mapping in the deep expansion network for realizing forward transformation and reverse transformation between an echo domain and an image domain, and obtaining an initial image of the deep expansion network by adopting the reverse RM mapping.
4. The method for single-bit millimeter wave imaging based on range migration enhanced deep expansion network according to claim 3, wherein S2 further comprises forward RM mapping in discrete wavenumber domain implementation Reverse RM mapping The expression is as follows: ; ; Wherein, the And (3) with Representing the two-dimensional fourier transform and its inverse respectively, Representing an element-by-element multiplication, Representing the forward compensation factor determined by the system parameters, Representing the input image domain data of the image, As a result of the conjugate compensation factor, , Represents the conjugation of the polymer and the polymer, Echo domain data representing an input; will simulate a single bit echo Inputting the reverse RM mapping to obtain an initial image: 。
5. The single-bit millimeter wave imaging method based on the range migration enhanced depth expansion network according to claim 1, wherein each stage of the depth expansion network is formed by serially connecting a symbol consistency updating unit and a space perception spurious elimination unit, and inter-stage parameters are shared or independently set.
6. The single-bit millimeter wave imaging method based on the range migration enhanced depth expansion network according to claim 1, wherein S3 comprises: converting the output image of one stage on the depth expansion network back to an echo domain, and performing a derivative approximation on a symbol function by using a hyperbolic tangent function to generate a predicted single-bit echo; Based on the predicted single-bit echo, constructing saturation suppression weights by using a tanh derivative relation, transmitting weighted residual errors back to an image domain through concomitant mapping to generate image update gradients, and performing gradient descent update on an output image at the previous stage by using the image update gradients to generate an intermediate image.
7. The single-bit millimeter wave imaging method based on the range migration enhanced depth expansion network according to claim 1, wherein S4 comprises the steps of carrying out convolution operation on the amplitude of an intermediate image by utilizing a learnable convolution layer initialized to be Gaussian kernel, extracting local energy density reflecting the target aggregation degree and generating a pixel-level self-adaptive threshold matrix, and applying MCP near-end mapping to the intermediate image pixel by pixel to obtain phase output.
8. The single-bit millimeter wave imaging method based on the range migration enhanced depth expansion network according to claim 7, wherein S4 comprises: from intermediate images Extracting an amplitude chart: ; For amplitude map Performing two-dimensional convolution, and adopting linear rectification function to ensure non-negative so as to generate local energy density : ; Wherein, the A two-dimensional convolution operation is represented, Is a convolution kernel initialized to A Gaussian kernel; According to local energy density Generating an adaptive threshold matrix at the pixel level For any pixel position Has the following formula: ; Wherein, the In order for the threshold reference to be a learnable, A small constant to prevent zero removal; For intermediate images Applying MCP near-end mapping pixel by pixel to obtain stage output ; For intermediate images Is input as complex value And taking the value of the corresponding position in the adaptive threshold matrix as a threshold value Order-making Definition of the definition (When ) And is also provided with The MCP proximal map is: ; Wherein, the Is a non-convexity control parameter.
9. The single-bit millimeter wave imaging method based on a range migration enhanced depth expansion network according to claim 1, wherein in S5, training the depth expansion network comprises: pairing samples The input depth is used to spread the network, In order to simulate the single-bit echo data, For corresponding to true value image Applying a random undersampling mask in a measurement dimension, wherein the sampling rate is randomly changed within a range of 30% -100%, and adopting a mean square error after amplitude normalization as a loss function, wherein the network output takes the amplitude to participate in supervision: ; Wherein, the Indexing samples within a lot B is the sample sampling batch, and the sample is taken from the sample, Indicating the first of the batches Single bit echo samples after the application of a random undersampling mask, Representation and representation The corresponding truth value intensity map label is displayed, Is the Frobenius norm, The representation takes the largest element of the amplitude matrix for normalization, A small constant to prevent zero removal; Minimizing with adaptive moment estimation optimizers by co-propagating joint update parameter sets Comprising step sizes of each stage Scaling factor of tanh Threshold benchmark Convolution kernel 。

Description

Single-bit millimeter wave imaging method based on range migration enhanced depth expansion network Technical Field The invention belongs to the technical field of radar signal processing and intelligent imaging, and particularly relates to a single-bit millimeter wave imaging method based on a range migration enhanced depth expansion network. Background Millimeter wave imaging has certain penetrating capacity, and the electromagnetic wave is non-ionizing radiation, so that the electromagnetic wave has small influence on human bodies under the condition of meeting the exposure limit value of related radiation, and is widely applied to scenes such as target detection, security inspection imaging, industrial detection and the like. With the increase of bandwidth and resolution, the echo data volume is rapidly increased, the quantization precision of the acquisition end, the storage and transmission bandwidth and the pressure of the back-end computing resource are synchronously increased, and the power consumption and the cost are increased, so that the engineering application of the low-cost, light-weight and real-time imaging system is affected. The single-bit quantization only reserves echo symbol information, can obviously reduce the complexity of an acquisition link and reduce the data volume, thereby relieving the storage and transmission pressure, and is an important technical path for realizing a low-cost millimeter wave imaging system. In order to improve the single-bit imaging effect, strategies such as time-varying threshold and jitter (DITHERING) quantization have been proposed, wherein the time-varying threshold can be changed through threshold time or observation, amplitude related information can be introduced to a certain extent, reconstruction accuracy can be improved, random disturbance is superimposed before quantization in jitter quantization, nonlinear distortion and spectrum aliasing can be relieved, and point-by-point preservation of a threshold sequence can be avoided. On the basis, compressed sensing (Compressed Sensing, CS) provides a theoretical basis for sparse recovery under low-bit or underdetermined observation, and the combination of dithering quantification and single-bit CS can improve imaging quality by means of sparse prior and optimal reconstruction. In recent years, deep learning has also been introduced into the millimeter wave imaging reconstruction field, namely one type of method adopts a general convolution network to carry out end-to-end mapping, and the other type of method uses a deep expansion thought to expand an iterative solution process into a trainable network hierarchical structure, so that the potential of trainable iterative process is shown in sparse reconstruction, and the method is the closest technical route to the invention at present. However, the existing method still has obvious defects when aiming at single-bit quantitative millimeter wave imaging. Firstly, single-bit quantization has extremely strong compression on amplitude information and is not conductive in a sign function, so that the traditional processing flow based on linear observation assumption is difficult to directly adapt, and is often represented by contrast reduction, detail loss and artifact increase, scattering intensity is easy to distort, and the target strength relationship is difficult to stably maintain. Secondly, the range migration effect is more obvious in millimeter wave imaging, the steps of frequency domain transformation, phase compensation, frequency domain resampling and the like related to the range migration are frequently required to be frequently executed in reconstruction, iteration cost is high, meanwhile, the influence of super parameters such as step length, regularization weight and the like on convergence speed and imaging quality is obvious, parameter adjustment cost is high, and accuracy and instantaneity are difficult to achieve. Thirdly, a general convolution network is adopted to carry out end-to-end mapping, the method is influenced by information deficiency and strong nonlinearity under the single-bit quantification condition, training and generalization are more difficult, stability is easy to be reduced when system parameters or working environments change, depth development special research for single-bit quantification millimeter wave imaging is still relatively insufficient, and systematic work is still lacking in the aspects of how to effectively characterize symbol observation characteristics, adapt range migration compensation and trainable iterative processes, press calculation complexity to an engineering acceptable range under the conditions of undistorted scattering strong and weak relation, controllable artifact and the like. Disclosure of Invention The invention provides a single-bit millimeter wave imaging method based on a range migration enhanced depth expansion network, which aims to solve the problems of amplitude information los