A method and system for adaptive image enhancement and turbulence removal

Through adaptive preprocessing and deep learning technology, combined with Fourier transform and sparse representation, dynamic filters and loss functions are constructed, which solves the problem of poor turbulence processing in the existing technology, significantly improving image clarity and fidelity , protect detailed information and adapt to turbulent interference in complex environments.

CN118552427BActive Publication Date: 2024-11-15CHANGSHA CHAOCHUANG ELECTRONICS TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410993030.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-24
Publication Date
2024-11-15
Estimated Expiration
2044-07-24

AI Technical Summary

Technical Problem

Existing deturbulence techniques have limited effects when dealing with complex and variable turbulence interference. Traditional methods cannot accurately recover details and often lead to loss of edge information, especially when dealing with images with large dynamic range and subtle texture features, and lack of them. Ability to dynamically adapt to turbulent characteristics.

Method used

Using image adaptive enhancement and deturbulence methods, combined with adaptive gamma correction, spatial adaptive mean denoising, Fourier transform, high-frequency energy screening, sparse dictionary construction and deep convolutional neural network (U-Net architecture) , construct dynamic filters and loss functions, perform fine preprocessing and deep learning training, significantly improve the clarity and fidelity of the image under turbulent conditions.

Benefits of technology

Effectively reduce the blur and distortion of the image by turbulence, significantly improve the sharpness, contrast and color fidelity of the image, protect important details and edge information, avoid excessive smoothing, improve the targeted and generalized processing effect, and ensure denoising The naturalness and accuracy of the results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118552427B_ABST
    Figure CN118552427B_ABST
Patent Text Reader

Abstract

The present invention discloses an image adaptive enhancement and de-turbulence method and system, the method comprising: S1: inputting a turbulence interference image as an original image, and preprocessing the original image to obtain a preprocessed image; S2: detecting turbulence features in the preprocessed image, and extracting the area affected by turbulence; S3: enhancing the area affected by turbulence; S4: constructing a de-turbulence network based on a deep convolutional neural network; S5: setting a loss function and training the de-turbulence network, and inputting the turbulence interference image into the trained de-turbulence network after processing according to steps S1 to S3 to obtain a de-turbulence result image. The present invention can significantly improve the clarity and fidelity of images under turbulent conditions through a series of sophisticated preprocessing steps and deep learning technology.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of turbulence removal, and in particular to an image adaptive enhancement and turbulence removal method and system. Background Art

[0002] In the field of image processing, especially in visual systems in complex environments such as aerial photography, remote monitoring and astronomical observation, image distortion caused by turbulence is a long-standing challenge. Turbulence effects can cause image blur and distortion, seriously affecting image quality and subsequent analysis and processing. Traditional denoising and image restoration methods, such as median filtering, bilateral filtering and classical deblurring algorithms, can improve image quality to a certain extent, but often fail to achieve satisfactory processing results when faced with complex and changeable turbulence interference. These methods either cannot accurately restore details or over-smoothing leads to loss of edge information, especially when processing images with large dynamic range and subtle texture features.

[0003] Most existing de-turbulence techniques are based on fixed filters or simple model assumptions, lacking the ability to dynamically adapt to turbulence characteristics, and therefore have limited effectiveness in dealing with nonlinear, non-steady-state turbulence interference in actual scenes. In addition, traditional image preprocessing steps, such as simple grayscale adjustment and noise removal, often ignore the diversity of image content and cannot optimize processing effects in a targeted manner. Summary of the invention

[0004] In view of this, the present invention provides an image adaptive enhancement and de-turbulence method, the purpose of which is to significantly improve the clarity and fidelity of images under turbulent conditions through a series of sophisticated preprocessing steps and deep learning technology.

[0005] To achieve the above object, the present invention provides an image adaptive enhancement and turbulence removal method, comprising the following steps:

[0006] S1: input the turbulence interference image as the original image, and preprocess the original image to obtain the preprocessed image;

[0007] S2: Detect turbulence features in the preprocessed image and extract the area affected by turbulence;

[0008] S3: Enhance the area affected by turbulence;

[0009] S4: Building a de-turbulence network based on deep convolutional neural network;

[0010] S5: setting a loss function and training the de-turbulence network, processing the turbulence interference image according to steps S1 to S3 and inputting it into the trained de-turbulence network to obtain a de-turbulence result image.

[0011] As a further improvement method of the present invention:

[0012] Optionally, the preprocessing includes adaptive gamma correction and spatial adaptive mean denoising, specifically:

[0013] S11: Adaptive gamma correction:

[0014] Calculate the original image Cumulative distribution histogram of :

[0015] ;

[0016] in, for The gray value is The number of pixels; is less than or equal to the gray value The number of pixel values ​​that appear; ;

[0017] Calculate gamma correction value :

[0018] ;

[0019] in, for In Location The value at and Corresponding to Centered The mean and median of the cumulative distribution histogram of pixels in the neighborhood;

[0020] Apply gamma correction value to obtain gamma corrected image , specifically:

[0021] ;

[0022] in, for In Location The value at for In Location The value at

[0023] S12: Spatial adaptive mean denoising:

[0024] Computing spatial similarity measures :

[0025] ;

[0026] in, For location As a benchmark, position Relative to Similarity measure of ; is a natural constant; for In Location The value at and They are grayscale bandwidth and spatial bandwidth respectively;

[0027] Calculate the weighted average denoising result :

[0028] ;

[0029] in, For location Centered on The pixel position within the neighborhood; for In Location The weighted average denoising result is the preprocessed image.

[0030] Optionally, the step S2 detects turbulence features in the preprocessed image and segments the area affected by turbulence, including:

[0031] S21: Fourier transform:

[0032] Perform a two-dimensional discrete Fourier transform on the preprocessed image to obtain a frequency domain image , specifically:

[0033] ;

[0034] in, is the frequency domain coordinate; for exist The value at is the two-dimensional discrete Fourier transform function;

[0035] S22: High frequency energy screening:

[0036] Define high frequency cutoff frequency , perform high-frequency energy screening on the frequency domain image to obtain a high-frequency energy map , specifically:

[0037] ;

[0038] in, for exist The value at

[0039] S23: Inverse Fourier transform and threshold segmentation:

[0040] Use the inverse Fourier transform to transform the high-frequency energy map back to the spatial domain to obtain the spatial domain energy map , and use the global threshold Segment the turbulent region, specifically:

[0041] ;

[0042] ;

[0043] in, for In Location The value at is the two-dimensional discrete Fourier inverse transform function; is the segmented area affected by turbulence.

[0044] Optionally, in step S3, enhancing the detected area affected by turbulence to reduce the influence of turbulence includes:

[0045] S31: Build a sparse dictionary:

[0046] Select the area not affected by turbulence in the preprocessed image and build a dictionary of pixel blocks ; Based on the sparsity theory, each pixel block in the area affected by turbulence is represented as a sparse matrix of dictionary elements;

[0047] S32: Sparse coding solution:

[0048] for For each pixel block in, solve the sparse vector so that the pixel block can be represented by the dictionary The best linear combination approximation is:

[0049] ;

[0050] in, for by Centered A pixel block composed of pixels within the range; for In the dictionary is a sparse vector based on

[0051] The solution is found by minimizing the following loss:

[0052] ;

[0053] in, is the minimum function; is the regularization parameter; Used to calculate the 2-norm; Used to calculate the 1 norm;

[0054] S33: Pixel Block Repair:

[0055] Using the sparse vector obtained by solution , from the dictionary Reconstruct the pixel blocks in the area affected by turbulence to obtain the repaired pixel blocks , and replace the corresponding pixel blocks in the preprocessed image to obtain an enhanced image .

[0056] Optionally, in step S4, constructing a de-turbulence network based on a deep convolutional neural network includes:

[0057] The deep convolutional neural network U-Net architecture is improved to obtain a de-turbulence network, wherein the improvement is a feature map obtained after the convolution operation in U-Net. Add dynamic filters, specifically:

[0058] S41: Dynamic filter generation:

[0059] A learnable Convolutional Layer Apply to Generate position-sensitive weights , specifically:

[0060] ;

[0061] in, , , and They are The height, width and number of channels, is the dynamic filter size;

[0062] Will Split into indivual dimensional vector, each vector Corresponds to A location in , and used to construct a The local filter , for dimensional real number set, specifically:

[0063] ;

[0064] in, for Middle vectors;

[0065] S42: Dynamic convolution operation:

[0066] for Each pixel position in , apply the corresponding dynamic filter to perform convolution operation, specifically:

[0067] ;

[0068] in, for In Location The filtering result of for In the location Centered on Position in window The value of for In Location The value of .

[0069] Optionally, in step S5, a loss function is set and a de-turbulence network is trained, and the turbulence interference image is processed according to steps S1 to S3 and input into the trained de-turbulence network to obtain a de-turbulence result image, including:

[0070] S51: Setting loss function:

[0071] Loss function of the de-turbulence network Loss of pixels , structural similarity index loss and blur loss Composition, specifically:

[0072] ;

[0073] in, , and is the weight of different losses;

[0074] S52: Optimize network parameters:

[0075] The parameters of the de-turbulence network are optimized based on stochastic gradient descent, specifically:

[0076] ;

[0077] in, and Separate Second and Parameters of the de-turbulence network at the time of update; is the learning rate; for about The partial derivative of

[0078] S53: Generate de-turbulence image:

[0079] The turbulence interference image is processed according to steps S1 to S3 and input into the de-turbulence network trained after iteration according to steps S51 and S52 to obtain a de-turbulence result image.

[0080] The present invention also discloses an image adaptive enhancement and turbulence removal system, comprising:

[0081] Preprocessing module: input turbulence interference image as original image and preprocess the original image;

[0082] Turbulence detection module: detects turbulence features in the preprocessed image and extracts the area affected by turbulence;

[0083] Enhancement module: enhances the area affected by turbulence;

[0084] Network building module: Building a de-turbulence network based on a deep convolutional neural network;

[0085] Network training and application module: set the loss function and train the de-turbulence network, process the turbulence interference image according to steps S1 to S3, and input it into the trained de-turbulence network to obtain the de-turbulence result image.

[0086] Beneficial effects:

[0087] Through the combination of adaptive preprocessing steps and deep learning technology, the present invention can effectively reduce the blur and distortion of images caused by turbulence, significantly improve the clarity, contrast and color fidelity of images, and make visual information more accurate and easy to identify.

[0088] The application of spatial adaptive mean denoising and sparse representation technology can accurately repair areas affected by turbulence, while protecting important details and edge information in the image and avoiding the over-smoothing problem common in traditional methods.

[0089] The introduction of dynamic filters enables the de-turbulence network to automatically optimize the processing strategy according to the specific features of the input image, which not only enhances the targeted processing effect, but also improves the processing speed and the generalization ability of the algorithm, enabling it to better adapt to turbulence interference in different scenarios and conditions.

[0090] Through the designed loss function, the present invention can balance pixel-level accuracy, structural similarity and image sharpness during the training process, ensuring that the output result of the turbulence removal network is closer to the real turbulence-free image, thereby improving the overall performance of the model and the naturalness of denoising. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 A schematic diagram of a process of an image adaptive enhancement and turbulence removal method according to an embodiment of the present invention;

[0092] Figure 2 Schematic diagram of turbulence removal results according to an embodiment of the present invention, wherein (a) is a turbulence interference image, and (b) is a turbulence removal result image. DETAILED DESCRIPTION

[0093] The present invention is further described below in conjunction with the accompanying drawings, but the present invention is not limited in any way. Any changes or substitutions made based on the teachings of the present invention belong to the protection scope of the present invention.

[0094] Embodiment 1:

[0095] An image adaptive enhancement and de-turbulence method, such as Figure 1 As shown, the following steps are included:

[0096] S1: Input the turbulence interference image as the original image, and preprocess the original image to obtain the preprocessed image:

[0097] The turbulence interference image is input as the original image, and the original image is preprocessed, and the preprocessing includes adaptive gamma correction and spatial adaptive mean denoising, specifically:

[0098] S11: Adaptive gamma correction:

[0099] Calculate the original image Cumulative distribution histogram of :

[0100] ;

[0101] in, for The gray value is The number of pixels; is less than or equal to the gray value The number of pixel values ​​that appear; ;

[0102] Calculate gamma correction value :

[0103] ;

[0104] in, for In Location The value at and Corresponding to Centered The mean and median of the cumulative distribution histogram of pixels in the neighborhood;

[0105] Apply gamma correction value to obtain gamma corrected image , specifically:

[0106] ;

[0107] in, for In Location The value at for In Location The value at

[0108] S12: Spatial adaptive mean denoising:

[0109] Computing spatial similarity measures :

[0110] ;

[0111] in, For location As a benchmark, position Relative to Similarity measure of ; is a natural constant; for In Location The value at and are grayscale bandwidth and spatial bandwidth respectively, which are 7 and 10 respectively in this embodiment;

[0112] Calculate the weighted average denoising result :

[0113] ;

[0114] in, For location Centered on The pixel position within the neighborhood; for In Location The weighted average denoising result is the preprocessed image.

[0115] Preprocessing removes or mitigates non-turbulence related noise in the image, providing a cleaner and easier to analyze base image for the subsequent de-turbulence algorithm. This helps the algorithm focus on solving the main problem - image distortion caused by turbulence, without being disturbed by other factors.

[0116] Through technologies such as adaptive gamma correction and spatial adaptive mean denoising, the brightness and contrast of the image can be effectively adjusted and noise can be reduced, making subsequent processing steps such as feature extraction and region segmentation more accurate and efficient. Good preprocessing can highlight the key information in the image and reduce the difficulty and complexity of algorithm processing.

[0117] S2: Detect the turbulence features in the preprocessed image and extract the area affected by turbulence:

[0118] S21: Fourier transform:

[0119] Perform a two-dimensional discrete Fourier transform on the preprocessed image to obtain a frequency domain image , specifically:

[0120] ;

[0121] in, is the frequency domain coordinate; for exist The value at is the two-dimensional discrete Fourier transform function;

[0122] S22: High frequency energy screening:

[0123] Define high frequency cutoff frequency In this embodiment , is the width of the preprocessed image; high-frequency energy is screened on the frequency domain image to obtain a high-frequency energy map , specifically:

[0124] ;

[0125] in, for exist The value at

[0126] S23: Inverse Fourier transform and threshold segmentation:

[0127] Use the inverse Fourier transform to transform the high-frequency energy map back to the spatial domain to obtain the spatial domain energy map , and use the global threshold To segment the turbulent region, in this embodiment, The mean value is:

[0128] ;

[0129] ;

[0130] in, for In Location The value at is the two-dimensional discrete Fourier inverse transform function; is the segmented area affected by turbulence.

[0131] By converting the image from the spatial domain to the frequency domain through Fourier transform, the low-frequency and high-frequency components in the image can be effectively separated. High-frequency energy screening is particularly targeted at turbulence effects in images, because turbulence often appears as obvious high-frequency signals in the frequency domain, which helps to directly and efficiently identify areas in the image affected by turbulence.

[0132] Compared with directly performing complex texture or edge detection in the spatial domain, processing in the frequency domain can achieve the screening of specific frequency features at a lower computational cost. , high-frequency components associated with turbulence can be quickly located and extracted, avoiding exhaustive pixel-level analysis of the entire image.

[0133] S3: Enhance the area affected by turbulence to obtain an enhanced image:

[0134] S31: Build a sparse dictionary:

[0135] Select the area not affected by turbulence in the preprocessed image and build a dictionary of pixel blocks ; Based on the sparsity theory, each pixel block in the area affected by turbulence is represented as a sparse matrix of dictionary elements;

[0136] S32: Sparse coding solution:

[0137] for For each pixel block in, solve the sparse vector so that the pixel block can be represented by the dictionary The best linear combination approximation is:

[0138] ;

[0139] in, for by Centered A pixel block composed of pixels within the range; for In the dictionary is a sparse vector based on

[0140] The solution is found by minimizing the following loss:

[0141] ;

[0142] in, is the minimum function; is the regularization parameter; Used to calculate the 2-norm; Used to calculate the 1 norm;

[0143] S33: Pixel Block Repair:

[0144] Using the sparse vector obtained by solution , from the dictionary Reconstruct the pixel blocks in the area affected by turbulence to obtain the repaired pixel blocks , and replace the corresponding pixel blocks in the preprocessed image to obtain an enhanced image .

[0145] Using sparse representation theory, we can find a few key elements that best represent the characteristics of the area affected by turbulence from a large number of dictionary elements. This method helps to accurately restore image details while removing noise, especially for areas with rich textures or complex structures.

[0146] dictionary It is constructed from areas not affected by turbulence, which means it contains the pure features of the image. This adaptively constructed dictionary can better match the specific image content, providing higher recovery flexibility and adaptability, and is more effective than fixed filters or templates.

[0147] S4: Building a turbulence removal network based on deep convolutional neural network:

[0148] The deep convolutional neural network U-Net architecture is improved to obtain a de-turbulence network, wherein the improvement is a feature map obtained after the convolution operation in U-Net. Add dynamic filters, specifically:

[0149] S41: Dynamic filter generation:

[0150] A learnable Convolutional Layer Apply to Generate position-sensitive weights , specifically:

[0151] ;

[0152] in, , , and They are The height, width and number of channels, is the dynamic filter size, which is 3 in this embodiment;

[0153] Will Split into indivual dimensional vector, each vector Corresponds to A location in , and used to construct a The local filter , for dimensional real number set, specifically:

[0154] ;

[0155] in, for Middle vectors;

[0156] S42: Dynamic convolution operation:

[0157] for Each pixel position in , apply the corresponding dynamic filter to perform convolution operation, specifically:

[0158] ;

[0159] in, for In Location The filtering result of for In the location Centered on Position in window The value of for In Location The value of .

[0160] In addition, the de-turbulence network is applied: the enhanced image is input into the de-turbulence network to obtain the de-turbulence image predicted by the de-turbulence network:

[0161]

[0162] in, For the de-turbulence network; The de-turbulenced image predicted by the de-turbulence network.

[0163] During the dynamic filter generation process, The position-sensitive weights learned by the convolutional layer can generate customized filters for each feature position. This not only enhances the network's ability to distinguish features at different positions, but also makes the filtering operation more targeted, improving the accuracy and effectiveness of feature expression.

[0164] The traditional convolution kernel is fixed on the entire image, while the dynamic filter generates different filtering parameters for different areas of the image, allowing the network to adaptively process the changing turbulence features in the image, especially in complex and changeable turbulence environments, improving denoising and fidelity.

[0165] S5: Set the loss function and train the de-turbulence network. After processing the turbulence interference image according to steps S1 to S3, input it into the trained de-turbulence network to obtain the de-turbulence result image:

[0166] S51: Setting loss function:

[0167] Loss function of the de-turbulence network Loss of pixels , structural similarity index loss and blur loss Composition, specifically:

[0168] ;

[0169] ;

[0170] ;

[0171] ;

[0172] in, It is a real turbulence-free image; Calculate the function for the structural similarity index; is the variance calculation function; and They are and The gradient of , and are weights of different losses, which are 1, 0.8 and 0.2 in this embodiment;

[0173] S52: Optimize network parameters:

[0174] The parameters of the de-turbulence network are optimized based on stochastic gradient descent, specifically:

[0175] ;

[0176] in, and Separate Second and Parameters of the de-turbulence network at the time of update; is the learning rate; for about The partial derivative of

[0177] S53: Generate de-turbulence image:

[0178] The turbulence interference image is processed according to steps S1 to S3 and then input into the de-turbulence network trained after iteration according to steps S51 and S52 to obtain the de-turbulence result image. Figure 2 Schematic diagram of the turbulence removal result, where (a) is the turbulence interference image and (b) is the turbulence removal result image.

[0179] Combining pixel loss, structural similarity index loss and blur loss can guide network learning from multiple dimensions. Pixel loss ensures the pixel-level proximity between the predicted image and the real image; SSIM loss promotes the similarity between the generated image and the target image in structure and texture; blur loss helps avoid over-smoothing of the predicted image and keeps the image details clear. This multi-objective optimization strategy is conducive to improving the quality and naturalness of de-turbulenced images.

[0180] Specifically, the above-mentioned image adaptive enhancement and turbulence removal method of this embodiment is used to Figure 2 The original image (turbulence interference image) shown in the left figure is processed to obtain the de-turbulence image (the result image of de-turbulence) as shown in Figure 2 As shown in the picture on the right.

[0181] Example 2

[0182] An image adaptive enhancement and de-turbulence system, including the following five modules:

[0183] Preprocessing module: input turbulence interference image as original image and preprocess the original image;

[0184] Turbulence detection module: detects turbulence features in the preprocessed image and extracts the area affected by turbulence;

[0185] Enhancement module: enhances the area affected by turbulence;

[0186] Network building module: Building a de-turbulence network based on a deep convolutional neural network;

[0187] Network training and application module: set the loss function and train the de-turbulence network, process the turbulence interference image according to steps S1 to S3, and input it into the trained de-turbulence network to obtain the de-turbulence result image.

[0188] It should be noted that the serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments. And the terms "including", "comprising" or any other variants thereof in this article are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "including a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0189] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0190] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. An image adaptive enhancement and turbulence removal method, characterized in that: The following steps are involved: S1: input the turbulence interference image as the original image, and preprocess the original image to obtain the preprocessed image; S2: Detect the turbulence features in the preprocessed image and extract the area affected by turbulence ; S3: Enhance the area affected by turbulence; The step S3 includes the following steps: S31: Build a sparse dictionary: Select the area not affected by turbulence in the preprocessed image and build a dictionary of pixel blocks ; Based on the sparsity theory, each pixel block in the area affected by turbulence is represented as a sparse matrix of dictionary elements; S32: Sparse coding solution: for For each pixel block in, solve the sparse vector so that the pixel block can be represented by the dictionary The best linear combination approximation is: ; in, for by Centered A pixel block composed of pixels within the range; for In the dictionary is a sparse vector based on ; The solution is found by minimizing the following loss: ; in, is the minimum function; is the regularization parameter; Used to calculate the 2-norm; Used to calculate the 1 norm; S33: Pixel Block Repair: Using the sparse vector obtained by solution , from the dictionary Reconstruct the pixel blocks in the area affected by turbulence to obtain the repaired pixel blocks , and replace the corresponding pixel blocks in the preprocessed image to obtain an enhanced image ; S4: Building a de-turbulence network based on deep convolutional neural network; The step S4 includes the following steps: The deep convolutional neural network U-Net architecture is improved to obtain a de-turbulence network, wherein the improvement is a feature map obtained after the convolution operation in U-Net Add dynamic filters, specifically: S41: Dynamic filter generation: A learnable Convolutional Layer Apply to Generate position-sensitive weights , specifically: ; in, , , and They are The height, width and number of channels, is the dynamic filter size; Will Split into indivual dimensional vector, each vector Corresponds to A location in , and used to construct a The local filter , for dimensional real number set, specifically: ; in, for Middle vectors; S42: Dynamic convolution operation: for Each pixel position in , apply the corresponding dynamic filter to perform convolution operation, specifically: ; in, for In Location The filtering result of for In the location As the center, Position in window The value of for In Location The value of S5: setting a loss function and training the de-turbulence network, processing the turbulence interference image according to steps S1 to S3 and inputting it into the trained de-turbulence network to obtain a de-turbulence result image.

2. The image adaptive enhancement and turbulence removal method according to claim 1, characterized in that: In step S1, the preprocessing includes adaptive gamma correction and spatial adaptive mean denoising, specifically: S11: Adaptive gamma correction: Calculate the original image Cumulative distribution histogram of : ; in, for The gray value is The number of pixels; is less than or equal to the gray value The number of pixel values ​​that appear; ; Calculate gamma correction value : ; in, for In Location The value at and Corresponding to Centered The mean and median of the cumulative distribution histogram of pixels in the neighborhood; Apply gamma correction value to obtain gamma corrected image , specifically: ; in, for In Location The value at for In Location The value at S12: Spatial adaptive mean denoising: Computing spatial similarity measures : ; in, For location As a benchmark, position Relative to Similarity measure of ; is a natural constant; for In Location The value at and They are grayscale bandwidth and spatial bandwidth respectively; Calculate the weighted average denoising result : ; in, For location As the center, The pixel position within the neighborhood; for In Location The weighted average denoising result is the preprocessed image.

3. The image adaptive enhancement and turbulence removal method according to claim 2, characterized in that: The step S2 includes the following steps: S21: Fourier transform: Perform a two-dimensional discrete Fourier transform on the preprocessed image to obtain a frequency domain image , specifically: ; in, is the frequency domain coordinate; for exist The value at is the two-dimensional discrete Fourier transform function; S22: High frequency energy screening: Define high frequency cutoff frequency , perform high-frequency energy screening on the frequency domain image to obtain a high-frequency energy map , specifically: ; in, for exist The value at S23: Inverse Fourier transform and threshold segmentation: Use the inverse Fourier transform to transform the high-frequency energy map back to the spatial domain to obtain the spatial domain energy map , and use the global threshold Segment the turbulent region, specifically: ; ; in, for In Location The value at is the two-dimensional discrete Fourier inverse transform function; is the segmented area affected by turbulence.

4. The image adaptive enhancement and turbulence removal method according to claim 3, characterized in that: The step S5 includes the following steps: S51: Setting loss function: Loss function of the de-turbulence network Loss of pixels , structural similarity index loss and blur loss Composition, specifically: ; in, , and is the weight of different losses; S52: Optimize network parameters: The parameters of the de-turbulence network are optimized based on stochastic gradient descent, specifically: ; in, and Separate Second and Parameters of the de-turbulence network at the time of update; is the learning rate; for about The partial derivative of S53: Generate de-turbulence image: The turbulence interference image is processed according to steps S1 to S3 and input into the de-turbulence network trained after iteration according to steps S51 and S52 to obtain a de-turbulence result image.

5. An image adaptive enhancement and turbulence removal system, characterized in that: include: Preprocessing module: input turbulence interference image as original image and preprocess the original image; Turbulence detection module: detects turbulence features in the preprocessed image and extracts the area affected by turbulence; Enhancement module: enhances the area affected by turbulence; Network building module: Building a de-turbulence network based on a deep convolutional neural network; Network training and application module: set the loss function and train the de-turbulence network. After the turbulence interference image is processed by the pre-processing module, turbulence detection module and enhancement module, it is input into the trained de-turbulence network to obtain the de-turbulence result image. To realize an image adaptive enhancement and de-turbulence method as described in any one of claims 1-4.