A wavefield-based method and apparatus for imaging deblurring

Abstract

The application provides a wave field-based imaging degradation elimination method and device. The method provided by the application comprises the following steps: acquiring an input image to be processed; obtaining a phase feature of each channel of the input image and a light wave propagation distance parameter of each channel by parallel prediction through a double-branch network; and estimating a global haze intensity of the input image in combination with prior information of the input image; constructing a virtual wave field in a complex number form according to a pixel intensity of the input image and the phase feature of each channel; converting the virtual wave field to a spatial frequency domain, constructing an angular spectrum transfer function according to the light wave propagation distance parameter of each channel and in combination with a light wave propagation physical rule, and inversely correcting phase distortion in the frequency domain in a channel-by-channel manner; adaptively generating a frequency domain amplitude gain function according to the global haze intensity, and compensating for the energy attenuated due to scattering in the frequency domain; inversely converting the wave field signal after the joint correction in the frequency domain back to a spatial image domain, fusing and refining the image features of multiple channels, and outputting a haze-free image.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method and apparatus for eliminating imaging degradation based on wave field. Background Technology

[0002] Image dehazing aims to recover clear scene radiation from observation images degraded by atmospheric scattering media (such as fog and haze). It is a key pre-processing technology in many fields and directly affects the accuracy and reliability of subsequent target recognition, scene understanding and decision control.

[0003] Current mainstream image dehazing methods can be divided into two main categories: one is traditional methods based on manual priors, such as dark channel priors, color attenuation priors, and nonlocal haze lines, which model and correct the fog distribution by constructing physical statistical models; the other is deep learning-based methods, from convolutional neural networks to Transformer architectures, which significantly improve the dehazing reconstruction quality in complex scenes through feature learning driven by massive amounts of data. However, existing image dehazing methods still have many problems: First, most are limited to pixel intensity domain operations, ignoring frequency domain distortion and multi-channel differential scattering characteristics at the wave optics level, and failing to fully explore frequency domain structural information; second, there is an irreconcilable contradiction between high-precision models and lightweight requirements, and large networks cannot meet the computing power requirements of low-latency scenarios such as real-time video and high frame rate streaming; third, existing lightweight dehazing networks lack strong physical prior constraints, which easily produce fog residue, structural blur, or ringing halo artifacts; fourth, a few frequency domain learning methods are still purely data-driven, without combining the physical laws of light wave propagation, lacking effective physical inductive biases, and limiting generalization ability.

[0004] Therefore, there is an urgent need for a method that can effectively adapt to various complex haze degradation scenarios while meeting the requirements of low computing power and low latency real-time deployment, and significantly improve the detail restoration and image quality stability of dehazed images. Summary of the Invention

[0005] In view of this, this application provides a wave field-based imaging degradation elimination method and apparatus, which can effectively adapt to various complex haze degradation scenarios while meeting the requirements of low computing power and low latency real-time deployment, and significantly improve the detail restoration and image quality stability of dehazed images.

[0006] Specifically, this application is implemented through the following technical solution:

[0007] The first aspect of this application provides a wavefield-based imaging degradation elimination method, the method comprising:

[0008] The input image to be processed is acquired, and the channel-wise phase features and channel-wise light wave propagation distance parameters of the input image are obtained through parallel prediction using a dual-branch network. The global haze intensity of the current input image is estimated by combining the prior information of the input image. The input image is an RGB hazy image.

[0009] Based on the pixel intensity of the input image and the channel-by-channel phase features, a virtual wavefield in complex form is constructed;

[0010] The virtual wave field is converted to the spatial frequency domain, and an angular spectrum transfer function is constructed based on the channel-by-channel optical wave propagation distance parameters and combined with the physical rules of optical wave propagation. The phase distortion in the frequency domain is then reversed for each channel.

[0011] Based on the global haze intensity, an adaptive frequency domain amplitude gain function is generated to compensate for the amplitude of energy attenuated by scattering in the frequency domain.

[0012] The wave field signal, after joint correction in the frequency domain, is inversely converted back to the spatial image domain. Multi-channel image features are then fused and refined to output a fog-free image.

[0013] A second aspect of this application provides an imaging degradation elimination device based on wave field, the device comprising a prediction module, a construction module, a processing module, and an output module;

[0014] The prediction module is used to acquire the input image to be processed, obtain the channel-by-channel phase features and channel-by-channel light wave propagation distance parameters of the input image through parallel prediction using a dual-branch network, and estimate the global haze intensity of the current input image by combining prior information of the input image; the input image is an RGB hazy image.

[0015] The construction module is used to construct a virtual wave field in complex form based on the pixel intensity of the input image and the channel-by-channel phase features;

[0016] The processing module is used to convert the virtual wave field to the spatial frequency domain, construct the angular spectrum transfer function based on the channel-by-channel optical wave propagation distance parameters and combined with the physical rules of optical wave propagation, and perform reverse correction of phase distortion in the frequency domain for each channel.

[0017] The processing module is also used to adaptively generate a frequency domain amplitude gain function based on the global haze intensity, and to compensate for the amplitude of the energy attenuated by scattering in the frequency domain.

[0018] The output module is used to inversely convert the wave field signal after frequency domain joint correction back to the spatial image domain, fuse and refine the multi-channel image features, and output a fog-free image.

[0019] The wavefield-based imaging degradation elimination method and apparatus provided in this application first predicts channel-by-channel phase features and channel-by-channel optical wave propagation distance parameters in parallel through a dual-branch network. Simultaneously, it estimates the global haze intensity by combining prior image information. This simultaneous acquisition of three key parameters efficiently matches the haze condition and optical propagation characteristics of a single RGB hazy image, providing accurate preliminary data for subsequent calculations. Next, a complex virtual wavefield is constructed based on the image pixel intensity and the predicted channel-by-channel phase features, transforming the two-dimensional image data into a wavefield expression that conforms to the laws of optical propagation. This achieves a reasonable conversion from image signal to optical wavefield, laying the foundation for frequency domain calculations. Finally, the virtual wavefield is converted to the spatial frequency domain and combined with light waves... The propagation physics rules and channel-by-channel light wave propagation distance parameters are used to construct an angular spectral transfer function to complete the inverse correction of frequency domain phase distortion. Then, based on the global haze intensity, a frequency domain amplitude gain function is adaptively generated to compensate for the energy attenuation caused by scattering. The frequency domain joint correction is completed simultaneously from both phase and amplitude dimensions, specifically addressing the phase distortion and energy attenuation problems caused by haze, and significantly improving the degradation elimination effect. Finally, the corrected wave field is inversely transformed back to the spatial image domain, and the image quality is optimized and the defects generated in the transformation process are corrected through multi-channel feature fusion and refinement. The final output is a haze-free image. The whole process relies on the optical propagation principle to ensure the physical rationality of the dehazing process and improves the adaptability to different haze scenarios. Attached Figure Description

[0020] Figure 1 A flowchart of the wave field-based imaging degradation elimination method provided in Embodiment 1 of this application;

[0021] Figure 2 This is a schematic diagram of the wave field-based imaging degradation elimination device provided in Embodiment 2 of this application. Detailed Implementation

[0022] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0023] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used herein are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0024] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0025] The following specific embodiments are given to illustrate the technical solution of this application in detail.

[0026] Figure 1 The flowchart illustrates the wavefield-based imaging degradation elimination method provided in Embodiment 1 of this application. Please refer to... Figure 1 The method provided in this embodiment may include:

[0027] S101. Obtain the input image to be processed, and obtain the channel-by-channel phase features and channel-by-channel light wave propagation distance parameters of the input image through parallel prediction using a dual-branch network. Combine the prior information of the input image to estimate the global haze intensity of the current input image.

[0028] The input image is an RGB image with fog.

[0029] Specifically, the input image refers to the RGB hazy image to be dehazed. It is a color image whose image quality has been degraded due to atmospheric scattering media such as fog and haze, and contains three color channels: red, green, and blue. The input image suffers visual degradation due to atmospheric scattering, which not only results in an overall decrease in pixel intensity, but also in different degradation of each color channel depending on the wavelength of light. At the same time, it is accompanied by phase distortion and loss of high-frequency details in the frequency domain.

[0030] A dual-branch network refers to a lightweight parallel network structure composed of two independent sub-networks, including a phase estimation sub-network and a distance prediction sub-network. Both sub-networks simultaneously receive the same input image and operate independently, outputting two different types of physical parameters in parallel. The channel-by-channel phase feature refers to the phase information obtained for each color channel of the RGB image. Since light scattering caused by haze alters the phase of light wave propagation, this feature characterizes the phase distortion state of light waves in each channel, reflecting potential information such as scene structure and edge direction obscured by haze. The channel-by-channel light wave propagation distance parameter corresponds to the parameters predicted for each color channel of the RGB image. Due to the influence of Mie scattering, Rayleigh scattering, and other effects, the propagation characteristics of light of different wavelengths differ in haze media. This parameter characterizes the equivalent propagation distance of light waves in each channel in the scattering medium, matching the differentiated degradation patterns of different channels. The global haze intensity refers to a scalar parameter statistically calculated based on prior image information, used to quantify the overall haze density of the entire image, intuitively reflecting the severity of the overall impact of haze scattering on the image.

[0031] In specific implementation, the input image's channel-wise phase features and channel-wise optical wave propagation distance parameters are obtained through parallel prediction using a dual-branch network. This includes: constructing a dual-branch network consisting of a phase estimation subnetwork and a distance prediction subnetwork; simultaneously inputting the input image into both the phase estimation subnetwork and the distance prediction subnetwork; the phase estimation subnetwork encoding the spatial structure and edge information of the input image and outputting the channel-wise phase features of each channel; the distance prediction subnetwork predicting the optical wave propagation distance parameters corresponding to each channel based on the global degradation degree of the input image; and the phase estimation subnetwork and the distance prediction subnetwork performing parallel operations.

[0032] Specifically, a dual-branch parallel network structure is pre-built, consisting of two independent and parallel sub-networks responsible for phase feature extraction and light wave propagation distance parameter prediction, respectively. A single RGB hazy input image is used as a unified input and simultaneously fed into both sub-networks. The phase estimation sub-network performs layer-by-layer feature encoding on the pixel spatial distribution, texture contours, and edge structure information of the input image, learning the haze distortion features of each color channel, and finally outputs independent phase features for each of the three RGB channels. The distance prediction sub-network scans the haze degradation level of the entire input image, captures the differential degradation features of different channels, and predicts the light wave propagation distance parameters adapted to each RGB color channel. The two sub-networks operate synchronously and infer in parallel throughout the process, without waiting for or interfering with each other, and simultaneously output complete channel-by-channel phase features and channel-by-channel light wave propagation distance parameters.

[0033] Optionally, the global haze intensity of the current input image is estimated by combining prior information of the input image, including: calculating the initial transmission features of the input image based on the dark channel prior to obtain an initial transmittance distribution that reflects the overall haze distribution of the input image; inputting the initial transmittance distribution into a feature perception network, which outputs residual compensation information for the initial transmission features; fusing the initial transmission features with the residual compensation information, and obtaining a range-normalized transmittance distribution through numerical truncation; and converting the global haze intensity based on the global statistical features of the transmittance distribution.

[0034] Specifically, dark channel features are extracted pixel-by-pixel from the entire input image to generate an initial transmissivity map for the entire image. This initial transmissivity map represents the basic distribution of fog in various parts of the input image. The obtained initial transmissivity map is used as input and fed into a pre-defined lightweight feature perception network (MLP). The network automatically learns and outputs a residual compensation feature map to correct the initial transmissivity deviation. The initial transmissivity map and the residual compensation feature map are added and fused pixel-by-pixel. Simultaneously, numerical truncation constraints are applied to the fused data to standardize the values ​​to a valid range, resulting in a normalized and accurate transmissivity distribution map. The global transmissivity mean of the entire accurate transmissivity distribution map is calculated as a global transmissivity scalar. Based on the complementary relationship between transmissivity and fog intensity (i.e., the sum of global transmissivity and global fog intensity is 1), a unique global scalar value is calculated using a fixed conversion formula; this value represents the global fog intensity of the input image.

[0035] For example, in one embodiment, the initial transmittance of the input image is quickly calculated using dark channel prior (DCP). It is fed into a very lightweight multilayer perceptron (MLP) to output residual compensation values. Adding the two together and truncating them gives the accurate transmittance. Then, a global smog intensity scalar was defined. .

[0036] It should be noted that the dual-branch network is pre-trained during prediction. The training process and training data for the dual-branch network are as follows. The first part of the training data is the RESIDE-ITS (Indoor Training Set) dataset. This dataset provides 13,990 pairs of high-quality synthetic indoor foggy images and their corresponding real fog-free (clear) image pairs. Furthermore, to comprehensively evaluate the generalization ability and real-time processing performance of this application under different scenes, resolutions, and fog concentrations, the following benchmark sets were used for performance verification: SOTS (Synthetic Objective Testing Set) benchmark set: used to comprehensively evaluate the model's basic defogging performance and color fidelity. Specifically, these include: SOTS-Indoor (containing 500 indoor foggy images), SOTS-Outdoor (containing 500 outdoor foggy images), SOTS-Mix (a mixture of the above indoor and outdoor test sets, totaling 1000 test images), and the RS-Haze high-resolution benchmark set (to further examine the model's ability to reconstruct structures and suppress halos when processing high-resolution images and dealing with extremely complex synthetic foggy scenes, the RS-Haze dataset was introduced. This test set contains 500 pairs of extremely challenging foggy / foggy image pairs, specifically designed to verify the robustness of advanced defogging network architectures under extreme conditions).

[0037] Optionally, the method further includes: calculating a fog concentration score for each frame from real images by analyzing the dark channel features and image contrast of video frames, and selecting clear images with fog concentration scores below a preset threshold; for the clear images, extracting single-channel Gaussian noise and performing a two-dimensional Gaussian blur operation to construct a spatially continuous and locally consistent fog layer, and adjusting the fog concentration by controlling the variance of the Gaussian noise and the size of the Gaussian blur kernel; linearly fusing the clear images with the white fog layer based on an atmospheric scattering model to generate synthetic fog images with different fog concentrations, which serve as a training dataset; and training a dual-branch network based on the training dataset.

[0038] Specifically, real-world video frames are captured frame by frame. The dark channel mean and global contrast of each frame are calculated. These two values ​​are then weighted and summed according to preset weights to obtain a single-frame fog concentration score. Video frames are sorted based on their fog concentration scores, and frames with scores below a preset threshold are selected as clear images. For the selected clear images, a single-channel Gaussian noise map is generated. A two-dimensional Gaussian operator is then used to blur the noise map, resulting in a continuously and uniformly distributed fog layer. By modifying the Gaussian noise variance and the Gaussian blur kernel size, fog layers of different concentrations are created. Following the linear fusion rules of the atmospheric scattering model, the clear images and corresponding fog layers are weighted and superimposed pixel by pixel to generate synthetic fogged images of different fog concentrations. All images are then compiled to form a training dataset.

[0039] For example, in one embodiment, the linear fusion process can be represented as:

[0040] ;

[0041] in, For synthesized fog images; Transmittance at each pixel location; For a clear image; The fog color is a constant (take white 255).

[0042] The method provided in this embodiment relies on real images to select clear images of real scenes as base samples. It combines dark channel and contrast indicators to quantify fog concentration and accurately select target scene images, ensuring that the base samples are consistent with actual application scenarios. Then, a fog layer is constructed by Gaussian noise and Gaussian blur, and the fog concentration is flexibly adjusted by noise variance and blur kernel size. This can generate simulated fog images with natural fog layer distribution and rich concentration gradients in batches. Finally, the linear fusion of images and fog layers is completed based on an atmospheric scattering model, so that the degradation law of the synthesized image matches the real atmospheric scattering characteristics. This not only greatly expands the number of samples and enriches the types of fog concentration scenes, but also makes the synthesized data highly close to the real scene in terms of fog state performance, color features, and degradation logic. Using this dataset to train a dual-branch network can effectively improve the network's adaptability and generalization performance to different fog concentrations and real scenes, while eliminating the cost of large-scale on-site collection of fog samples, thus improving the efficiency and practicality of dataset production.

[0043] S102. Based on the pixel intensity of the input image and the channel-by-channel phase characteristics, construct a virtual wave field in complex form.

[0044] Specifically, pixel intensity refers to the brightness value corresponding to each pixel in the input image, with brightness and darkness information recorded for each location in separate channels. Virtual wave field refers to a simulated light wave field obtained by combining the pixel intensity and phase characteristics of each channel of the image in a complex format, used to simulate the light wave pattern of light propagating in a medium.

[0045] In specific implementation, the pixel intensity of each channel of the input image is processed by amplitude conversion to obtain the initial amplitude of the virtual wave field of each channel; the initial amplitude of the virtual wave field of each channel is fused with the channel-by-channel phase feature of the corresponding channel in a complex form to generate the complex virtual wave field of each channel; the real part and imaginary part of the complex virtual wave field are obtained by trigonometric function operation of the initial amplitude and the channel-by-channel phase feature, respectively.

[0046] Specifically, the initial amplitude of the virtual wave field for each channel is obtained by taking the square root of the pixel intensity value of each color channel of the input image. The initial amplitude of each channel is then subjected to trigonometric function operations with the phase characteristics of the corresponding channel to calculate the cosine and sine components. The complex virtual wave field for each channel is constructed by using the cosine component as the real part of the complex number and the sine component as the imaginary part of the complex number.

[0047] For example, in one embodiment, the virtual wave field can be represented as:

[0048] ;

[0049] in, For virtual wave fields; The initial amplitude of the virtual wavefield for each channel; This represents the phase characteristics of each channel.

[0050] S103. The virtual wave field is converted to the spatial frequency domain. Based on the channel-by-channel optical wave propagation distance parameters and combined with the physical rules of optical wave propagation, the angular spectrum transfer function is constructed, and the phase distortion in the frequency domain is reversed for each channel.

[0051] Specifically, the spatial frequency domain is a mathematical representation of image signals. Through Fourier transform, the virtual wavefield is converted from a spatial domain with pixel position (x, y) as the variable to a frequency domain with spatial frequency (wavenumber) as the variable, facilitating subsequent phase correction calculations based on the laws of light wave propagation. The physical laws governing light wave propagation refer to the physical laws followed by angular spectrum propagation theory. The core principle is that when light waves propagate in a homogeneous medium, plane waves with different spatial frequency components evolve independently according to their respective propagation directions, and their phase changes with propagation distance satisfy a fixed physical relationship. The angular spectrum transfer function is a frequency domain operator constructed based on the physical laws of light wave propagation and combined with channel-by-channel light wave propagation distance parameters. It describes the phase change law of different spatial frequency components during propagation and its function is to apply inverse phase compensation to the frequency domain virtual wavefield, thereby correcting phase distortion.

[0052] In practice, the virtual wavefields of each channel are preprocessed by boundary windowing to suppress ringing artifacts generated during frequency domain transformation. The preprocessed virtual wavefields of each channel are then mapped to the corresponding spatial frequency domain through Fourier transform. For each channel, the optical wave propagation distance parameter corresponding to the channel is used, combined with the preset wavelength parameter that conforms to the optical propagation characteristics, and the reverse propagation angular spectrum transfer function is constructed channel by channel according to the physical mechanism of angular spectrum propagation. In the frequency domain, for the virtual wavefields of each channel, invalid wave components that exceed the effective propagation range of the haze medium are filtered out. Using the reverse propagation angular spectrum transfer function of each channel, the frequency domain phase distortion is reverse-corrected channel by channel to restore the phase distribution characteristics of each channel when the optical wave is not affected by haze scattering interference.

[0053] Specifically, a window function is applied to the edge region of the virtual wavefield of each channel to perform boundary smoothing; a two-dimensional fast Fourier transform is performed on the windowed virtual wavefield of each channel to map it from the spatial domain to the spatial frequency domain; the wavenumber corresponding to each channel is calculated according to a preset wavelength, and then the phase compensation factor of each frequency component is calculated by substituting it into the angular spectrum propagation formula in combination with the optical wave propagation distance parameter of the corresponding channel, thereby constructing the inverse propagation angular spectrum transfer function; in the frequency domain, high-frequency invalid wave components are filtered out according to a set threshold; the virtual wavefield of each channel in the frequency domain is multiplied point by point with the inverse propagation angular spectrum transfer function of the corresponding channel to complete the inverse correction of the frequency domain phase distortion, and the corrected frequency domain wavefield of each channel is obtained.

[0054] S104. Generate a frequency domain amplitude gain function adaptively based on the global haze intensity to compensate for the amplitude of energy attenuated by scattering in the frequency domain.

[0055] Specifically, the frequency domain amplitude gain function is a frequency domain operator generated based on the global haze intensity. It is used to adaptively compensate for the energy attenuated by haze scattering in the frequency domain wavefield. Taking the global haze intensity as input, the frequency domain amplitude gain function generates gain coefficients related to spatial frequencies, adjusting the amplitude of each component of the frequency domain wavefield. This restores the high-frequency and overall energy weakened by scattering, bringing the wavefield amplitude back to a level unaffected by haze interference.

[0056] In specific implementation, the global haze intensity is input into a preset linear transformation layer, and feature mapping is performed through a Sigmoid activation function to obtain gain control parameters. The linear transformation layer includes learnable weight parameters and bias parameters, and the output of the Sigmoid activation function is scaled within a preset numerical range. The gain control parameters are limited to a preset minimum and maximum value. A parameterized Sigmoid-type frequency domain amplitude gain function is constructed using the gain control parameters as adjustment variables. The frequency domain amplitude gain function uses spatial frequency as the independent variable. The gain value in the first region is stably maintained at a preset baseline level, while the gain value in the second region smoothly transitions to a preset upper limit level as the frequency increases. The spatial frequency of the first region is lower than that of the second region. According to the frequency domain energy attenuation characteristics of each channel, the corresponding frequency domain amplitude gain function is matched channel by channel to adaptively repair the frequency domain amplitude under different haze concentrations.

[0057] Specifically, the global haze intensity scalar is input into the linear transformation layer, multiplied by the weight parameter matrix, and then a bias parameter is added to obtain the linear transformation result. This result is then input into the Sigmoid activation function to obtain an intermediate value between 0 and 1. This intermediate value is linearly scaled according to preset minimum and maximum values ​​to obtain gain control parameters within a defined range. Using the scaled gain control parameters as adjustment variables, a frequency domain amplitude gain function is constructed. First, a base gain value and a high-frequency gain upper limit are set. Then, the gain inflection point frequency and curve steepness coefficient are set to ensure that the constructed frequency domain amplitude gain function stably maintains the base gain in the region below the inflection point frequency, and smoothly transitions to the high-frequency gain upper limit as the frequency increases in the region above the inflection point frequency. For each channel, a corresponding frequency domain amplitude gain function is matched according to its frequency domain energy attenuation characteristics. The wave field amplitude in the frequency domain is multiplied point by point to complete adaptive compensation of frequency domain amplitude under different haze concentrations.

[0058] For example, in one embodiment, a set of parameters is defined to control the shape of the frequency domain gain. ,in, It is a set of adjustable parameter vectors used to control the shape of the frequency domain gain curve; This is the base gain value for low frequencies; This represents the gain variation amplitude from low frequency to high frequency; The radius of the inflection point frequency of the gain change; This is the kurtosis coefficient of the gain curve.

[0059] Based on the overall smog intensity The gain parameter corresponding to the current frame is dynamically calculated through linear mapping:

[0060] ;

[0061] in, Use the Sigmoid activation function; This is the gain parameter; This represents the minimum value of the gain parameter; This represents the maximum value of the gain parameter; These are learnable weighting coefficients used to adjust the degree of influence of global smog intensity; The global haze intensity of the input image; This is a bias term.

[0062] Furthermore, in terms of frequency radius Construct a parameterized Sigmoid-type frequency domain amplitude-gain function with the variable as the independent variable. :

[0063] ;

[0064] in, It is the frequency domain amplitude gain function; This is the base gain value for low frequencies; This represents the high-frequency gain limit. The radius of the inflection point frequency of the gain change; This is the kurtosis coefficient of the gain curve; For frequency. This function ensures low and stable gain in the low-frequency range (i.e., overall illumination and color), while at the cutoff frequency... Controlled nonlinear amplification is applied to the high-frequency bands (i.e., textures and edges).

[0065] S105. The wave field signal after frequency domain joint correction is inversely converted back to the spatial image domain, and the multi-channel image features are fused and refined to output a fog-free image.

[0066] Specifically, the spatial image domain is the original image representation using pixel positions (x, y) as coordinates, distinct from the frequency domain, and is an image format presented as a two-dimensional pixel array. Multi-channel image features refer to the image signal features corresponding to the R / G / B color channels obtained after the inverse transformation of the virtual wavefield of each channel after correction. These features contain the corrected amplitude and phase information of each channel.

[0067] In specific implementation, the wavefield signal after frequency domain joint correction is inversely converted back to the spatial image domain. This includes: in the frequency domain, performing element-wise complex fusion operations on the virtual wavefield, angular spectral transfer function, and frequency domain amplitude gain function of each channel to obtain the modulated frequency domain wavefield signal; performing a two-dimensional inverse Fourier transform on the modulated frequency domain wavefield signal of each channel to restore the frequency domain complex wavefield to the spatial domain complex form; and calculating the squared modulus of the spatial domain complex wavefield signal of each channel to extract the preliminary dehazing pixel intensity features of the corresponding channel.

[0068] Specifically, in the frequency domain, for each channel's virtual wavefield, angular spectral transfer function, and frequency domain amplitude gain function, complex multiplication is performed point-by-point according to pixel position to obtain the modulated frequency domain wavefield signal; for each channel's modulated frequency domain wavefield signal, a two-dimensional inverse fast Fourier transform is performed to restore the frequency domain complex wavefield to the spatial domain complex wavefield; for each channel's spatial domain complex wavefield, its modulus is first calculated, and then the modulus result is squared to obtain the preliminary dehazing pixel intensity characteristics of the corresponding channel.

[0069] Optionally, the multi-channel image features are fused and refined to output a haze-free image, including: stitching together the preliminary dehazing pixel intensity features of each channel along the channel dimension to form a complete multi-channel feature tensor; inputting the multi-channel feature tensor into a decoding network, and eliminating grid artifacts introduced by frequency domain transformation through cross-channel feature fusion, local detail enhancement, and color alignment processing, while repairing image edge texture information; mapping the final three-channel pixel feature map from the output of the decoding network, and outputting a haze-free image after numerical range normalization processing.

[0070] Specifically, the initial dehazing pixel intensity features of each channel are sequentially concatenated into a three-dimensional multi-channel feature tensor along the channel dimension. This feature tensor is then fed into the decoding network, where cross-channel information fusion is achieved through convolutional layers. This enhances the details of local areas, aligns the color components of each channel, eliminates grid artifacts introduced by frequency domain transformation, and restores the texture information of image edges. The decoding network outputs pixel feature maps of the three channels, and their numerical range is then normalized to the standard image range to obtain the final haze-free image.

[0071] For example, in one embodiment, the foggy wave field signal is... Angular spectral transfer function and frequency domain amplitude gain function Element-wise complex multiplication is performed to complete spectral phase correction and amplitude compensation.

[0072] ;

[0073] in, This is the wavefield signal after joint correction in the frequency domain; It is a signal with a foggy wave field; Angular spectrum transfer function; This is the frequency domain amplitude gain function.

[0074] A two-dimensional inverse fast Fourier transform (2D-IFFT) is performed on the modulated frequency domain signal to restore it to the spatial domain. The square of the modulus of the complex wave field is calculated to extract the intensity of the single-channel pixels after preliminary dehazing.

[0075] ;

[0076] in, The intensity of a single-channel pixel after initial dehazing; This is the wave field signal after frequency domain joint correction.

[0077] The reconstructed R, G, and B channels are concatenated along the channel dimension. This concatenation is then fed into a hybrid decoder consisting of alternating stacks of 1x1 convolutions, 3x3 depthwise separable convolutions, group normalization (GroupNorm), and SiLU activation functions. Through cross-channel information fusion and detail refinement within the local receptive field, minute grid artifacts that may remain due to frequency domain discrete transformation are eliminated, resulting in a final high-fidelity, fog-free image.

[0078] The method provided in this embodiment, in its first aspect, firstly predicts the channel-by-channel phase features and channel-by-channel optical wave propagation distance parameters in parallel through a dual-branch network, and simultaneously estimates the global haze intensity by combining prior image information. This simultaneous acquisition of three key parameters efficiently matches the haze conditions and optical propagation characteristics of a single RGB hazy image, providing accurate preliminary data for subsequent calculations. Secondly, based on the image pixel intensity and the predicted channel-by-channel phase features, a complex virtual wavefield is constructed, transforming the two-dimensional image data into a wavefield expression that conforms to the laws of optical propagation, achieving a reasonable conversion from image signal to optical wavefield, laying the foundation for frequency domain calculations. Finally, the virtual wavefield is converted to the spatial frequency domain, combined with the physics of light wave propagation... The angular spectral transfer function is constructed using the rules and channel-by-channel light wave propagation distance parameters to complete the inverse correction of frequency domain phase distortion. Then, based on the global haze intensity, a frequency domain amplitude gain function is adaptively generated to compensate for the energy attenuation caused by scattering. The frequency domain joint correction is completed simultaneously from both phase and amplitude dimensions, specifically addressing the phase distortion and energy attenuation problems caused by haze, and significantly improving the degradation elimination effect. Finally, the corrected wave field is inversely transformed back to the spatial image domain, and the image quality is optimized and the defects generated in the transformation process are corrected through multi-channel feature fusion and refinement. The final output is a haze-free image. The entire process relies on the optical propagation principle to ensure the physical rationality of the dehazing process and improves the adaptability to different haze scenarios.

[0079] Secondly, the virtual wavefield of each channel is first preprocessed by windowing at the boundary to suppress ringing artifacts caused by frequency domain transformation. Then, the wavefield is mapped to the spatial frequency domain through Fourier transform. Based on the preset optical wavelength parameters and the exclusive light wave propagation distance parameters of each channel, the reverse propagation angular spectrum transfer function is constructed for each channel in combination with the physical mechanism of angular spectrum propagation. At the same time, invalid wave components that exceed the effective propagation range of the haze medium are filtered out in the frequency domain, and the transfer function is used to complete the reverse correction of phase distortion for each channel. This method abandons the limitation of the traditional method of only adjusting grayscale in the pixel spatial domain. It strictly follows the physical law of light wave propagation to realize the differential phase inverse restoration of each channel. It can accurately correct the phase shift and distortion of each channel caused by haze scattering, effectively restore the original phase distribution of light waves, and avoid the problems of blurred image edges, structural distortion and artifact residue after dehazing.

[0080] Thirdly, the global haze intensity is connected to a linear transformation layer with learnable weights and biases. After Sigmoid activation mapping and numerical range scaling, a gain control parameter with a limited range is obtained. Then, a parameterized Sigmoid-type frequency domain amplitude gain function is constructed using this parameter as an adjustment variable. Based on spatial frequency, high and low frequency regions are distinguished and adaptive amplitude compensation is performed for each channel according to the frequency domain energy attenuation characteristics. The gain curve shape can be dynamically adjusted according to the haze concentration of different images. The low frequency region stably retains the basic brightness and color characteristics of the image, while the high frequency region smoothly enhances the detail energy attenuated by haze scattering. This can avoid the problems of over-enhancement or insufficient compensation caused by uniform gain, and can also specifically repair the frequency domain energy loss of each channel, greatly improving the ability to restore texture details and color fidelity of dehazed images.

[0081] Corresponding to the aforementioned embodiment of an imaging degradation elimination method based on wave field, this application also provides an embodiment of an imaging degradation elimination device based on wave field.

[0082] Figure 2 This is a schematic diagram of the wavefield-based imaging degradation elimination device provided in Embodiment 2 of this application. Please refer to... Figure 2 The apparatus provided in this embodiment includes a prediction module 210, a construction module 220, a processing module 230, and an output module 240.

[0083] The prediction module 210 is used to acquire the input image to be processed, obtain the channel-by-channel phase features and channel-by-channel light wave propagation distance parameters of the input image through parallel prediction using a dual-branch network, and estimate the global haze intensity of the current input image by combining prior information of the input image; the input image is an RGB hazy image.

[0084] The construction module 220 is used to construct a virtual wave field in complex form based on the pixel intensity of the input image and the channel-by-channel phase features;

[0085] The processing module 230 is used to convert the virtual wave field to the spatial frequency domain, construct the angular spectrum transfer function based on the channel-by-channel optical wave propagation distance parameters and combined with the physical rules of optical wave propagation, and perform reverse correction of phase distortion in the frequency domain by channel;

[0086] The processing module 230 is also used to adaptively generate a frequency domain amplitude gain function based on the global haze intensity, and to compensate for the amplitude of the energy attenuated by scattering in the frequency domain.

[0087] The output module 240 is used to inversely convert the wave field signal after frequency domain joint correction back to the spatial image domain, fuse and refine the multi-channel image features, and output a fog-free image.

[0088] The apparatus of this embodiment can be used to perform... Figure 1 The steps of the method embodiment shown are similar in principle and process, and will not be repeated here.

[0089] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0090] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and 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 application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0091] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A wavefield-based imaging degradation elimination method, characterized in that, The method includes: The input image to be processed is acquired, and the channel-wise phase features and channel-wise light wave propagation distance parameters of the input image are obtained through parallel prediction using a dual-branch network. The global haze intensity of the current input image is estimated by combining the prior information of the input image. The input image is an RGB hazy image. Based on the pixel intensity of the input image and the channel-by-channel phase features, a virtual wavefield in complex form is constructed; The virtual wave field is converted to the spatial frequency domain, and an angular spectrum transfer function is constructed based on the channel-by-channel optical wave propagation distance parameters and combined with the physical rules of optical wave propagation. The phase distortion in the frequency domain is then reversed for each channel. Based on the global haze intensity, an adaptive frequency domain amplitude gain function is generated to compensate for the amplitude of energy attenuated by scattering in the frequency domain. The wave field signal after frequency domain joint correction is inversely converted back to the spatial image domain, and the multi-channel image features are fused and refined to output a fog-free image. The method of obtaining the channel-by-channel phase features and channel-by-channel light wave propagation distance parameters of the input image through parallel prediction using a dual-branch network includes: Construct a two-branch network consisting of a phase estimation subnetwork and a distance prediction subnetwork; The input image is simultaneously input into the phase estimation subnetwork and the distance prediction subnetwork; The phase estimation subnetwork encodes the spatial structure and edge information of the input image and outputs the channel-by-channel phase features of each channel. The distance prediction subnetwork predicts the light wave propagation distance parameters corresponding to each channel based on the global degradation degree of the input image; the phase estimation subnetwork operates in parallel with the distance prediction subnetwork.

2. The method according to claim 1, characterized in that, The virtual wave field is converted to the spatial frequency domain. Based on the channel-by-channel optical wave propagation distance parameters and combined with the physical rules of optical wave propagation, an angular spectral transfer function is constructed. Inverse correction of phase distortion in the frequency domain is performed channel by channel, including: Each channel's virtual wavefield is preprocessed with boundary windowing to suppress ringing artifacts generated during frequency domain transformation. The preprocessed virtual wavefields of each channel are mapped to the corresponding spatial frequency domain through Fourier transform; For each channel, the optical wave propagation distance parameter corresponding to the channel is adopted, and combined with the preset wavelength parameter that conforms to the optical propagation characteristics, the reverse propagation angular spectrum transfer function is constructed channel by channel according to the physical mechanism of angular spectrum propagation; In the frequency domain, for each channel's virtual wave field, invalid wave components that exceed the effective propagation range of the haze medium are filtered out. Using the reverse propagation angular spectral transfer function of each channel, the frequency domain phase distortion is reverse-corrected channel by channel to restore the phase distribution characteristics of each channel's light waves when they are not affected by haze scattering interference.

3. The method according to claim 1, characterized in that, Based on the global haze intensity, an adaptive frequency domain amplitude gain function is generated to compensate for the energy attenuated by scattering in the frequency domain, including: The global haze intensity is input into a preset linear transformation layer, and feature mapping is performed through the Sigmoid activation function to obtain the gain control parameters. The linear transformation layer contains learnable weight parameters and bias parameters, and the output of the Sigmoid activation function is scaled within a preset numerical range. The gain control parameters are limited to a preset minimum and maximum value. Using the gain control parameters as adjustment variables, a parameterized Sigmoid-type frequency domain amplitude gain function is constructed; the frequency domain amplitude gain function uses the frequency radius as the independent variable, the gain value in the first region is stably maintained at a preset base level, the gain value in the second region smoothly transitions to a preset upper limit level as the frequency radius increases, and the frequency radius of the first region is smaller than that of the second region. Based on the frequency domain energy attenuation characteristics of each channel, the corresponding frequency domain amplitude gain function is matched for each channel to adaptively repair the frequency domain amplitude under different haze concentrations.

4. The method according to claim 1, characterized in that, The global haze intensity of the current input image is estimated by combining prior information of the input image, including: The initial transmittance features of the input image are calculated based on the dark channel prior, and the initial transmittance distribution reflecting the overall fog distribution of the input image is obtained. The initial transmittance distribution is input into a feature sensing network, which then outputs residual compensation information for the initial transmittance features. The initial transmission characteristics are fused with the residual compensation information, and a range-normalized transmittance distribution is obtained through numerical truncation processing. Based on the global statistical characteristics of the transmittance distribution, the global haze intensity is calculated.

5. The method according to claim 1, characterized in that, The wavefield signal, after frequency domain joint correction, is inversely converted back to the spatial image domain, including: In the frequency domain, the virtual wave field, angular spectrum transfer function, and frequency domain amplitude gain function of each channel are fused element by element to obtain the modulated frequency domain wave field signal. Two-dimensional inverse Fourier transform is performed on the frequency domain wave field signals modulated by each channel to restore the frequency domain complex wave field to the spatial domain complex form; The squared modulus value is calculated for the spatial domain complex wave field signal of each channel, and the preliminary dehazing pixel intensity features of the corresponding channel are extracted.

6. The method according to claim 1, characterized in that, Multi-channel image features are fused and refined to output a haze-free image, including: The initial dehazed pixel intensity features of each channel are concatenated along the channel dimension to form a complete multi-channel feature tensor. The multi-channel feature tensor is input into the decoding network, and through cross-channel feature fusion, local detail enhancement and color alignment processing, the grid artifacts introduced by frequency domain transformation are eliminated, and the image edge texture information is repaired at the same time. The final three-channel pixel feature map is generated by mapping the output of the decoding network. After numerical range normalization, a fog-free image is output.

7. The method according to claim 1, characterized in that, The method further includes: From real images, the fog density score of each frame is calculated by analyzing the dark channel features and image contrast of the video frames, and clear images with fog density scores below a preset threshold are selected. For the clear image, single-channel Gaussian noise is extracted and a two-dimensional Gaussian blur operation is performed to construct a spatially continuous and locally consistent fog layer. The fog concentration is adjusted by controlling the variance of the Gaussian noise and the size of the Gaussian blur kernel. Based on the atmospheric scattering model, the clear image is linearly fused with the white fog layer to generate synthetic fog images with different fog concentrations, which serve as the training dataset. A dual-branch network is trained based on the training dataset.

8. The method according to claim 1, characterized in that, Based on the pixel intensity of the input image and the channel-by-channel phase features, a virtual wavefield in complex form is constructed, including: The amplitude of the pixel intensity of each channel of the input image is converted to obtain the initial amplitude of the virtual wave field of each channel. The initial amplitude of the virtual wave field of each channel is fused with the channel-by-channel phase characteristics in a complex form to generate the complex virtual wave field of each channel; the real part and the imaginary part of the complex virtual wave field are obtained by trigonometric function operations on the initial amplitude and the channel-by-channel phase characteristics, respectively.

9. A wavefield-based imaging degradation cancellation device, characterized in that, The device includes a prediction module, a construction module, a processing module, and an output module; The prediction module is used to acquire the input image to be processed, obtain the channel-by-channel phase features and channel-by-channel light wave propagation distance parameters of the input image through parallel prediction using a dual-branch network, and estimate the global haze intensity of the current input image by combining prior information of the input image; the input image is an RGB hazy image. The construction module is used to construct a virtual wave field in complex form based on the pixel intensity of the input image and the channel-by-channel phase features; The processing module is used to convert the virtual wave field to the spatial frequency domain, construct the angular spectrum transfer function based on the channel-by-channel optical wave propagation distance parameters and combined with the physical rules of optical wave propagation, and perform reverse correction of phase distortion in the frequency domain for each channel. The processing module is also used to adaptively generate a frequency domain amplitude gain function based on the global haze intensity, and to compensate for the amplitude of the energy attenuated by scattering in the frequency domain. The output module is used to inversely convert the wave field signal after frequency domain joint correction back to the spatial image domain, fuse and refine the multi-channel image features, and output a fog-free image. The method of obtaining the channel-by-channel phase features and channel-by-channel light wave propagation distance parameters of the input image through parallel prediction using a dual-branch network includes: Construct a two-branch network consisting of a phase estimation subnetwork and a distance prediction subnetwork; The input image is simultaneously input into the phase estimation subnetwork and the distance prediction subnetwork; The phase estimation subnetwork encodes the spatial structure and edge information of the input image and outputs the channel-by-channel phase features of each channel. The distance prediction subnetwork predicts the light wave propagation distance parameters corresponding to each channel based on the global degradation degree of the input image; the phase estimation subnetwork operates in parallel with the distance prediction subnetwork.

Images