Remote sensing data intelligent noise reduction and enhancement system and method based on deep learning

Abstract

The invention discloses a remote sensing data intelligent noise reduction and enhancement system and method based on deep learning, and belongs to the technical field of image processing and data analysis. The method comprises the following steps: S1, data acquisition: acquiring a remote sensing image of a target ocean area through a satellite and an unmanned aerial vehicle sensor; s2, building an intelligent noise reduction model, carrying out feature training through a deep learning network, and carrying out feature noise reduction processing on the obtained remote sensing image data; s3, carrying out image registration and combination on the images subjected to feature noise reduction; s4, performing image post-processing on the image after image registration and merging; and S5, result visual display and data analysis feedback are carried out, and the processed image is visually displayed through image processing software. The deep learning method is used for training, summarized noise features are directly learned, training complexity caused by excessive noise is avoided, resources are greatly saved, and training time is shortened.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing and data analysis, and specifically to a remote sensing data intelligent noise reduction and enhancement system and method based on deep learning. Background Art

[0002] Image intelligent noise reduction technology is a technology that uses artificial intelligence and deep learning algorithms to reduce noise and interference in images and improve image quality.

[0003] In remote sensing image processing, especially the processing of remote sensing image noise, it has a very large impact on the image presentation and quality. Currently, Gaussian filtering, mean filtering and bilateral filtering are mostly used for noise processing of remote sensing images. If the image blur is caused by Gaussian noise in a single image, then using Gaussian filtering for denoising will have a better effect. However, in many cases, there is more than one image and more than one type of noise, and the image noise may be caused by multiple situations. The current machine learning-based denoising models rely on image data sets for self-learning, and summarize various noise characteristics through a large amount of image data. In this case, it takes a long time to train with a large number of images, resulting in a waste of resources and inability to be applied on a large scale. Moreover, the denoising model will be complicated to train due to excessive noise during the training process, resulting in poor denoising effect. Summary of the invention

[0004] The purpose of the present invention is to provide a remote sensing data intelligent noise reduction and enhancement system and method based on deep learning to solve the problems raised in the above background technology.

[0005] In order to solve the above technical problems, the present invention provides the following technical solution: a remote sensing data intelligent noise reduction and enhancement method based on deep learning, the method comprising the following steps:

[0006] S1. Data acquisition: collect remote sensing images of different bands of the target ocean area through satellite and drone sensors, including visible light images, near infrared images and short-wave infrared images;

[0007] S2. Build an intelligent noise reduction model, perform feature training through a deep learning network, and perform feature noise reduction processing on the acquired remote sensing image data to achieve intelligent noise reduction;

[0008] S3, performing image registration and merging on images of different bands after feature noise reduction to enhance the detail display of remote sensing images;

[0009] S4, performing image post-processing on the image after image registration and merging, including color correction and image enhancement, so as to reduce the chromatic aberration of the remote sensing image and realize image enhancement;

[0010] S5. Visual display of results and data analysis feedback: The processed images are visualized through image display software to facilitate user analysis and interpretation; the image data is analyzed in combination with specific application scenarios, and the analysis results are fed back to the intelligent noise reduction model to optimize the model.

[0011] The image registration and merging includes image registration, radiation correction and normalization processing, wherein image registration is to spatially register images of different bands so that they are aligned in the same coordinate system to ensure that the pixels of images of different bands correspond one to one; radiation correction is to perform radiation correction on the image to eliminate the radiation differences introduced by the sensor and atmospheric factors; and normalization processing is to standardize the pixel values ​​to the same range.

[0012] In step S1,

[0013] Satellite sensors have a wide coverage area and can provide multi-spectral high-resolution data; they are suitable for long-term monitoring of large areas; drone sensors have adjustable flight altitudes and high flexibility, which are suitable for detailed monitoring of local areas; they are suitable for high-precision monitoring and rapid response in specific areas;

[0014] The visible light image reflects the color and reflection characteristics of the ocean surface and is used to detect ocean surface pollution and algae coverage; the near-infrared image reflects the infrared reflection characteristics of vegetation and water bodies and is used to monitor the distribution and changes of marine vegetation and plankton; the short-wave infrared image has strong penetration and can provide information about water bodies and wetlands, and is used to monitor changes in ocean temperature, humidity and chemical composition.

[0015] In step S2, the feature training includes statistical analysis features, frequency domain analysis features, autocorrelation features and histogram features;

[0016] Statistical analysis features include:

[0017] Mean:

[0018]

[0019] Among them, Mean represents the mean value of pixels in the image; K represents the total number of pixels in the image, z i Represents the pixel value of point i in the image;

[0020] Variance:

[0021]

[0022] Among them, Variance represents the variance of image pixel values;

[0023] Standard Deviation:

[0024]

[0025] Among them, Standard Deviation represents the standard deviation of image pixel values;

[0026] Kurtosis Kur:

[0027]

[0028] Among them, Kur represents the kurtosis of the image;

[0029] The kurtosis reflects the sharpness of the image brightness distribution;

[0030] Frequency domain analysis feature: extract the frequency information of the image by performing frequency domain transformation on the image;

[0031] Fourier transform: converts an image from the spatial domain to the frequency domain and analyzes the frequency components of the image;

[0032]

[0033] Among them, F(u,v) represents the image in the frequency domain; f(x,y) represents the image in the spatial domain; M represents the size of the image in the x direction; N represents the size of the image in the y direction; u represents the horizontal frequency component in the frequency domain; v represents the vertical frequency component in the frequency domain; x and y represent the pixel coordinates in the spatial domain;

[0034] Power spectral density E: describes the energy distribution of different frequency components in the image;

[0035] E(u,v)=|F(u,v)| 2

[0036] Among them, E(u,v) represents the power of each frequency component in the frequency domain; |F(u,v)| represents the amplitude of each frequency component in the frequency domain;

[0037] The high-frequency and low-frequency components of the image are extracted through a bandpass filter;

[0038] High frequency component: satisfies u≥u c , v ≥ v c ,u c 1 / 2 of the horizontal frequency range; v c 1 / 2 of the vertical frequency range;

[0039] High frequency component energy E high :

[0040] Total energy E total :

[0041] High frequency component energy ratio HF:

[0042]

[0043] The high-frequency component and the low-frequency component of the image reflect the edge and texture features of the image and the overall smoothness of the image;

[0044] Autocorrelation feature: The autocorrelation function is used to describe the correlation between image pixels;

[0045]

[0046] Where K represents the total number of pixels in the image; R(Δx, Δy) represents the autocorrelation value between pixels at the displacement (Δx, Δy); f(x+Δx, y+Δy) represents the image after displacement;

[0047] Histogram features:

[0048] Histogram mean H Mean :

[0049]

[0050] Among them, H Mean represents the mean of the histogram image; L represents the number of gray levels; p(t) represents the probability of gray level t;

[0051] Histogram variance H var :

[0052]

[0053] Among them, H var represents the histogram variance;

[0054] Perform feature training on the intelligent denoising model.

[0055] In step S2, the noise category C k Including Gaussian noise, impulse noise, Poisson noise and shot noise, the specific judgment method is:

[0056] Select kurtosis Kur as a parameter in the statistical characteristics:

[0057]

[0058] Where Kur represents kurtosis; Indicates noise type C k The average value of the lower kurtosis; Indicates noise type C k The standard deviation under

[0059] Select the energy ratio of high-frequency components in the frequency domain features:

[0060]

[0061] Wherein, HF represents the high frequency component energy ratio; Indicates noise type C k The average value of the energy ratio of the lower high frequency components; Indicates noise type C k The standard deviation of the energy ratio of the lower high frequency components;

[0062] Select the zero-delay autocorrelation value as a parameter in the autocorrelation feature: set the values ​​of Δx and Δy to 0 to obtain the zero-delay autocorrelation value;

[0063]

[0064] Where AC represents the zero-delay autocorrelation value; Indicates noise type C k The average of the zero-lag autocorrelation values ​​under ; Indicates noise type C k The standard deviation of the zero-lag autocorrelation values ​​under ;

[0065] In the histogram feature, the histogram mean is selected as the parameter:

[0066]

[0067] Among them, H var represents the histogram variance; Indicates noise type C k The average value of the histogram variance under ; Indicates noise type C k The standard deviation of the histogram variance under ;

[0068] Calculate the probability:

[0069]

[0070] Among them, P(Kur), P(HF), P(AC) and P(H var ) represents the marginal probability of the feature; P(C k ) indicates noise type C k The prior probability of not observing the noise feature; P(Kur|C k ) indicates that in noise type C k The probability density function of kurtosis is: P(HF|C k ) indicates that in noise type C k The probability density function of the high-frequency component energy ratio is: P(AC|C k ) indicates that in noise type C kThe probability density function of the zero-delay autocorrelation value under var |C k ) indicates that in noise type C k Below, the probability density function of the histogram mean;

[0071]

[0072] in, Indicates the post-verification noise type, including Gaussian noise, impulse noise, Poisson noise, and shot noise;

[0073] By calculating the posterior probability of each noise type, and selecting the noise type with the largest posterior probability.

[0074] In step S2, the Gaussian noise is denoised using wavelet transform, and the specific steps are as follows:

[0075] S5-1, decompose the image into different scales and directions through wavelet transform, and perform threshold processing on the wavelet coefficients;

[0076] S5-2, performing wavelet transform on the image to obtain wavelet coefficients; performing soft threshold and hard threshold processing on the wavelet coefficients;

[0077] S5-3. Perform inverse wavelet transform on the processed wavelet coefficients to reconstruct the image.

[0078] In step S3, the impulse noise is denoised using adaptive median filtering, the specific steps are:

[0079] S6-1, initialize the window size;

[0080] S6-2, calculating the median, maximum and minimum values ​​of the pixels in the window;

[0081] S6-3, adjusting the window size according to whether the pixel value is between the median and the extreme value;

[0082] S6-4. Replace the central pixel with the median value of the adaptive window.

[0083] In step S3, the method for processing Poisson noise is to use Anscombe transform to transform the image containing Poisson noise into a Gaussian noise-like image, then use wavelet transform to reduce noise, and then perform inverse transform.

[0084] The Anscombe transformation is specifically:

[0085]

[0086] Among them, I' represents a Gaussian noise image; I represents a Poisson noise image;

[0087] The inverse transform is:

[0088]

[0089] Shot noise is reduced using gamma filtering.

[0090] In step S3, the image registration and merging adopts weighted average merging, specifically:

[0091]

[0092] in, Represents the pixel value of the merged image at (x, y); I a (x,y),I b (x,y) and I c (x, y) represents the pixel value of the visible light image, near infrared image and short wave infrared image at (x, y) respectively; w1, w2 and w3 represent the weights of the visible light image, near infrared image and short wave infrared image respectively; the weights are set according to the conclusions drawn from the experiment and the amount of information contained in the image, specifically w1 = 0.4, w2 = 0.3 and w3 = 0.3.

[0093] In step S4, the color correction and image enhancement are specifically as follows:

[0094] Color correction: adjust the gain of each channel to make the color of the image more balanced; select the white area in the image and perform white balance correction to ensure that the white area looks more natural; use the reference image to perform color mapping to make the color distribution of the target image consistent with the reference image;

[0095] Image enhancement: Enhance the contrast of the image to make the light and dark areas more distinct; perform unsharp mask processing to enhance the edges and details of the image to make the image clearer; convert to HSV color space to adjust color saturation and hue.

[0096] A remote sensing data intelligent denoising and enhancement system based on deep learning, which executes the remote sensing data intelligent denoising and enhancement method based on deep learning as claimed in claim 1, and the system includes a data acquisition module, an intelligent denoising model module, an image registration and merging module, an image post-processing module and a result display module;

[0097] The data acquisition module is used to collect remote sensing image data of the target area through satellite and UAV sensors; the intelligent noise reduction model module is used to process the collected remote sensing images, remove various types of noise, including Gaussian noise, impulse noise, Poisson noise and shot noise, and enhance the image quality; the image registration and merging module is used to register and merge images of multi-spectral bands to generate a comprehensive image containing more information and details; the image post-processing module is used to further optimize the image quality, enhance the details and contrast of the image, and perform color correction and image enhancement; the result display module is used to display the processed image to the user, provide visualization tools and interfaces, and support the user to further analyze and operate the image;

[0098] The output end of the data acquisition module is connected to the input end of the intelligent denoising model module; the output end of the intelligent denoising model module is connected to the input end of the image registration and merging module; the output end of the image registration and merging module is connected to the input end of the image post-processing module; the output end of the image post-processing module is connected to the input end of the result display module.

[0099] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0100] 1. The present invention analyzes a large number of remote sensing images after artificially collecting them to obtain noise characteristics of various images. The model summarizes the characteristics through a large amount of image data, and directly learns the summarized characteristics. There will be no excessive noise that complicates the training, which greatly saves resources and shortens the training time.

[0101] 2. The present invention performs different processing on different noises, and can achieve better noise reduction processing effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0102] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0103] Figure 1 It is a schematic diagram of the steps of the remote sensing data intelligent noise reduction and enhancement method based on deep learning of the present invention;

[0104] Figure 2 is the Gaussian noise histogram;

[0105] Figure 3 is the impulse noise histogram;

[0106] Figure 4 is the Poisson noise histogram;

[0107] Figure 5 is the shot noise histogram. DETAILED DESCRIPTION

[0108] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0109] See also Figure 1-Figure 5 The present invention provides a technical solution: a remote sensing data intelligent noise reduction and enhancement method based on deep learning, the method comprising the following steps:

[0110] S1. Data acquisition: collect remote sensing images of different bands of the target ocean area through satellite and drone sensors, including visible light images, near infrared images and short-wave infrared images;

[0111] S2. Build an intelligent noise reduction model, perform feature training through a deep learning network, and perform feature noise reduction processing on the acquired remote sensing image data to achieve intelligent noise reduction;

[0112] S3, performing image registration and merging on images of different bands after feature noise reduction to enhance the detail display of remote sensing images;

[0113] S4, performing image post-processing on the image after image registration and merging, including color correction and image enhancement, so as to reduce the chromatic aberration of the remote sensing image and realize image enhancement;

[0114] S5. Visual display of results and data analysis feedback: The processed images are visualized through image display software to facilitate user analysis and interpretation; the image data is analyzed in combination with specific application scenarios, and the analysis results are fed back to the intelligent noise reduction model to optimize the model.

[0115] The image registration and merging includes image registration, radiation correction and normalization processing, wherein image registration is to spatially register images of different bands so that they are aligned in the same coordinate system to ensure that the pixels of images of different bands correspond one to one; radiation correction is to perform radiation correction on the image to eliminate the radiation differences introduced by the sensor and atmospheric factors; and normalization processing is to standardize the pixel values ​​to the same range.

[0116] In step S1,

[0117] Satellite sensors have a wide coverage area and can provide multi-spectral high-resolution data; they are suitable for long-term monitoring of large areas; drone sensors have adjustable flight altitudes and high flexibility, which are suitable for detailed monitoring of local areas; they are suitable for high-precision monitoring and rapid response in specific areas;

[0118] The visible light image reflects the color and reflection characteristics of the ocean surface and is used to detect ocean surface pollution and algae coverage; the near-infrared image reflects the infrared reflection characteristics of vegetation and water bodies and is used to monitor the distribution and changes of marine vegetation and plankton; the short-wave infrared image has strong penetration and can provide information about water bodies and wetlands, and is used to monitor changes in ocean temperature, humidity and chemical composition.

[0119] In step S2, the feature training includes statistical analysis features, frequency domain analysis features, autocorrelation features and histogram features;

[0120] Statistical analysis features include:

[0121] Mean:

[0122]

[0123] Among them, Mean represents the mean value of pixels in the image; K represents the total number of pixels in the image, z i Represents the pixel value of point i in the image;

[0124] Variance:

[0125]

[0126] Among them, Variance represents the variance of image pixel values;

[0127] Standard Deviation:

[0128]

[0129] Among them, Standard Deviation represents the standard deviation of image pixel values;

[0130] Kurtosis Kur:

[0131]

[0132] Among them, Kur represents the kurtosis of the image;

[0133] The kurtosis reflects the sharpness of the image brightness distribution;

[0134] Frequency domain analysis feature: extract the frequency information of the image by performing frequency domain transformation on the image;

[0135] Fourier transform: converts an image from the spatial domain to the frequency domain and analyzes the frequency components of the image;

[0136]

[0137] Among them, F(u,v) represents the image in the frequency domain; f(x,y) represents the image in the spatial domain; M represents the size of the image in the x direction; N represents the size of the image in the y direction; u represents the horizontal frequency component in the frequency domain; v represents the vertical frequency component in the frequency domain; x and y represent the pixel coordinates in the spatial domain;

[0138] Power spectral density E: describes the energy distribution of different frequency components in the image;

[0139] E(u,v)=|F(u,v)| 2

[0140] Among them, E(u,v) represents the power of each frequency component in the frequency domain; |F(u,v)| represents the amplitude of each frequency component in the frequency domain;

[0141] The high-frequency and low-frequency components of the image are extracted through a bandpass filter;

[0142] High frequency component: satisfies u≥u c , v ≥ v c ,u c 1 / 2 of the horizontal frequency range; v c 1 / 2 of the vertical frequency range;

[0143] High frequency component energy E high :

[0144] Total energy E total :

[0145] High frequency component energy ratio HF:

[0146]

[0147] The high-frequency component and the low-frequency component of the image reflect the edge and texture features of the image and the overall smoothness of the image;

[0148] Autocorrelation feature: The autocorrelation function is used to describe the correlation between image pixels;

[0149]

[0150] Where K represents the total number of pixels in the image; R(Δx, Δy) represents the autocorrelation value between pixels at the displacement (Δx, Δy); f(x+Δx, y+Δy) represents the image after displacement;

[0151] Histogram features:

[0152] Histogram mean H Mean :

[0153]

[0154] Among them, H Mean represents the mean of the histogram image; L represents the number of gray levels; p(t) represents the probability of gray level t;

[0155] Histogram variance H var :

[0156]

[0157] Among them, H var represents the histogram variance;

[0158] Perform feature training on the intelligent denoising model.

[0159] In step S2, the noise category C k Including Gaussian noise, impulse noise, Poisson noise and shot noise, the specific judgment method is:

[0160] Select kurtosis Kur as a parameter in the statistical characteristics:

[0161]

[0162] Where Kur represents kurtosis; Indicates noise type C k The average value of the lower kurtosis; Indicates noise type C k The standard deviation under

[0163] Select the energy ratio of high-frequency components in the frequency domain features:

[0164]

[0165] Wherein, HF represents the high frequency component energy ratio; Indicates noise type C k The average value of the energy ratio of the lower high frequency components; Indicates noise type C k The standard deviation of the energy ratio of the lower high frequency components;

[0166] Select the zero-delay autocorrelation value as a parameter in the autocorrelation feature: set the values ​​of Δx and Δy to 0 to obtain the zero-delay autocorrelation value;

[0167]

[0168] Where AC represents the zero-delay autocorrelation value; Indicates noise type C k The average of the zero-lag autocorrelation values ​​under ; Indicates noise type C kThe standard deviation of the zero-lag autocorrelation values ​​under ;

[0169] In the histogram feature, the histogram mean is selected as the parameter:

[0170]

[0171] Among them, H var represents the histogram variance; Indicates noise type C k The average value of the histogram variance under ; Indicates noise type C k The standard deviation of the histogram variance under ;

[0172] Calculate the probability:

[0173]

[0174]

[0175] Among them, P(Kur), P(HF), P(AC) and P(H var ) represents the marginal probability of the feature; P(C k ) indicates noise type C k The prior probability of not observing the noise feature; P(Kur|C k ) indicates that in noise type C k The probability density function of kurtosis is: P9HF|C k ) indicates that in noise type C k The probability density function of the high-frequency component energy ratio is: P(AC|C k ) indicates that in noise type C k The probability density function of the zero-delay autocorrelation value under var |C k ) indicates that in noise type C k Below, the probability density function of the histogram mean;

[0176]

[0177] in, Indicates the post-verification noise type, including Gaussian noise, impulse noise, Poisson noise, and shot noise;

[0178] By calculating the posterior probability of each noise type, and selecting the noise type with the largest posterior probability.

[0179] In step S2, the Gaussian noise is denoised using wavelet transform, and the specific steps are as follows:

[0180] S5-1, decompose the image into different scales and directions through wavelet transform, and perform threshold processing on the wavelet coefficients;

[0181] S5-2, performing wavelet transform on the image to obtain wavelet coefficients; performing soft threshold and hard threshold processing on the wavelet coefficients;

[0182] S5-3. Perform inverse wavelet transform on the processed wavelet coefficients to reconstruct the image.

[0183] In step S3, the impulse noise is denoised using adaptive median filtering, the specific steps are:

[0184] S6-1, initialize the window size;

[0185] S6-2, calculating the median, maximum and minimum values ​​of the pixels in the window;

[0186] S6-3, adjusting the window size according to whether the pixel value is between the median and the extreme value;

[0187] S6-4. Replace the central pixel with the median value of the adaptive window.

[0188] In step S3, the method for processing Poisson noise is to use Anscombe transform to transform the image containing Poisson noise into a Gaussian noise-like image, then use wavelet transform to reduce noise, and then perform inverse transform.

[0189] The Anscombe transformation is specifically:

[0190]

[0191] Among them, I' represents a Gaussian noise image; I represents a Poisson noise image;

[0192] The inverse transform is:

[0193]

[0194] Shot noise is reduced using gamma filtering.

[0195] In step S3, the image registration and merging adopts weighted average merging, specifically:

[0196]

[0197] in, Represents the pixel value of the merged image at (x, y); I a (x,y),I b (x,y) and I c(x, y) represents the pixel value of the visible light image, near infrared image and short wave infrared image at (x, y) respectively; w1, w2 and w3 represent the weights of the visible light image, near infrared image and short wave infrared image respectively; the weights are set according to the conclusions drawn from the experiment and the amount of information contained in the image, specifically w1 = 0.4, w2 = 0.3 and w3 = 0.3.

[0198] In step S4, the color correction and image enhancement are specifically as follows:

[0199] Color correction: adjust the gain of each channel to make the color of the image more balanced; select the white area in the image and perform white balance correction to ensure that the white area looks more natural; use the reference image to perform color mapping to make the color distribution of the target image consistent with the reference image;

[0200] Image enhancement: Enhance the contrast of the image to make the light and dark areas more distinct; perform unsharp mask processing to enhance the edges and details of the image to make the image clearer; convert to HSV color space to adjust color saturation and hue.

[0201] A remote sensing data intelligent denoising and enhancement system based on deep learning, which executes the remote sensing data intelligent denoising and enhancement method based on deep learning as claimed in claim 1, and the system includes a data acquisition module, an intelligent denoising model module, an image registration and merging module, an image post-processing module and a result display module;

[0202] The data acquisition module is used to collect remote sensing image data of the target area through satellite and UAV sensors; the intelligent noise reduction model module is used to process the collected remote sensing images, remove various types of noise, including Gaussian noise, impulse noise, Poisson noise and shot noise, and enhance the image quality; the image registration and merging module is used to register and merge images of multi-spectral bands to generate a comprehensive image containing more information and details; the image post-processing module is used to further optimize the image quality, enhance the details and contrast of the image, and perform color correction and image enhancement; the result display module is used to display the processed image to the user, provide visualization tools and interfaces, and support the user to further analyze and operate the image;

[0203] The output end of the data acquisition module is connected to the input end of the intelligent denoising model module; the output end of the intelligent denoising model module is connected to the input end of the image registration and merging module; the output end of the image registration and merging module is connected to the input end of the image post-processing module; the output end of the image post-processing module is connected to the input end of the result display module.

[0204] In an embodiment of the present invention,

[0205] Step S1: Data acquisition

[0206] Data collection method:

[0207] Satellite sensors: Use satellite sensors to cover the target ocean area and obtain multi-spectral high-resolution data, including visible light imagery, near-infrared imagery, and short-wave infrared imagery.

[0208] Drone sensors: Use drones to conduct detailed monitoring in local areas and obtain high-precision visible light, near-infrared, and short-wave infrared images.

[0209] Data Type:

[0210] Visible light images: used to detect ocean surface pollution and algae cover.

[0211] Near-infrared images: used to monitor the distribution and changes of marine vegetation and plankton.

[0212] Shortwave infrared imagery: used to monitor changes in ocean temperature, humidity, and chemical composition.

[0213] Step S2: Build an intelligent noise reduction model

[0214] Feature training:

[0215] Statistical analysis features: Calculate the mean, variance, standard deviation, kurtosis, and skewness of an image.

[0216] Frequency domain analysis features: Extract the frequency components of the image through Fourier transform, and calculate the power spectrum density and high-frequency component energy ratio.

[0217] Autocorrelation feature: Calculates zero-delay autocorrelation value to describe the correlation between pixels.

[0218] Histogram features: Calculate the histogram mean and variance and analyze the image grayscale distribution.

[0219] Intelligent noise reduction model:

[0220] Gaussian noise: Use wavelet denoising to remove Gaussian noise through wavelet transform and threshold processing.

[0221] Pulse noise: Use adaptive median filtering to adjust the filter window size according to the noise density to remove pulse noise.

[0222] Poisson noise: Use Anscombe transform to convert Poisson noise into Gaussian noise, and then use wavelet denoising.

[0223] Shot Noise: Use gamma filtering to reduce noise.

[0224] Step S3: Image registration and merging

[0225] Image registration: spatially register images of different bands so that they are aligned in the same coordinate system to ensure one-to-one correspondence between pixels.

[0226] Radiometric Correction: Removes radiation differences introduced by the sensor and atmospheric factors.

[0227] Normalization: Standardize pixel values ​​to the same range.

[0228] Weighted average merging: Merge visible, near-infrared, and short-wave infrared images using weights determined experimentally.

[0229]

[0230] in, Represents the pixel value of the merged image at (x, y); I a (x,y),I b (x,y) and I c (x, y) represents the pixel value of the visible light image, near infrared image and short wave infrared image at (x, y) respectively; w1, w2 and w3 represent the weights of the visible light image, near infrared image and short wave infrared image respectively; the weights are set according to the conclusions drawn from the experiment and the amount of information contained in the image, specifically w1 = 0.4, w2 = 0.3 and w3 = 0.3.

[0231] Step S4: Image post-processing

[0232] Color Correction:

[0233] White balance correction: Select a white area in the image for white balance correction.

[0234] Color mapping: Use the reference image for color mapping so that the color distribution of the target image is consistent with the reference image.

[0235] Image Enhancement:

[0236] Contrast Enhancement: Increases the contrast between light and dark areas of an image.

[0237] Unsharp Masking: Enhances edges and details in images.

[0238] Color saturation and hue adjustment: Convert to HSV color space and adjust color saturation and hue Step S5: Result visualization and data analysis feedback

[0239] Visual display: Use image display software to display processed images to facilitate user analysis and interpretation.

[0240] Data analysis feedback: Analyze image data based on specific application scenarios, and feed the analysis results back to the intelligent denoising model to optimize the model.

[0241] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device 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, method, article or device.

[0242] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A remote sensing data intelligent noise reduction and enhancement method based on deep learning, characterized by: The method comprises the following steps: S1. Data acquisition: collect remote sensing images of different bands of the target ocean area through satellite and drone sensors, including visible light images, near infrared images and short-wave infrared images; S2. Build an intelligent noise reduction model, perform feature training through a deep learning network, and perform feature noise reduction on the acquired remote sensing image data; S3, performing image registration and merging on images of different bands after feature noise reduction; S4, performing image post-processing on the image after image registration and merging, including color correction and image enhancement; S5. Visual display of results and data analysis feedback: visualize the processed images through image display software; analyze the image data in combination with specific application scenarios, feed the analysis results back to the intelligent denoising model to optimize the model.

2. The remote sensing data intelligent noise reduction and enhancement method based on deep learning according to claim 1 is characterized in that: The image registration and merging includes image registration, radiation correction and normalization processing, wherein image registration is to spatially register images of different bands so that they are aligned in the same coordinate system to ensure that the pixels of images of different bands correspond one to one; radiation correction is to perform radiation correction on the image to eliminate the radiation differences introduced by the sensor and atmospheric factors; and normalization processing is to standardize the pixel values ​​to the same range.

3. The remote sensing data intelligent noise reduction and enhancement method based on deep learning according to claim 1 is characterized in that: In step S2, the feature training includes statistical analysis features, frequency domain analysis features, autocorrelation features and histogram features; Statistical analysis features include: Mean: Among them, Mean represents the mean value of pixels in the image; K represents the total number of pixels in the image, z i Represents the pixel value of point i in the image; Variance: Among them, Variance represents the variance of image pixel values; Standard Deviation: Among them, Standard Deviation represents the standard deviation of image pixel values; Kurtosis Kur: Among them, Kur represents the kurtosis of the image; The kurtosis reflects the sharpness of the image brightness distribution; Frequency domain analysis feature: extract the frequency information of the image by performing frequency domain transformation on the image; Fourier transform: converts an image from the spatial domain to the frequency domain and analyzes the frequency components of the image; Among them, F(u,v) represents the image in the frequency domain; f(x,y) represents the image in the spatial domain; M represents the size of the image in the x direction; N represents the size of the image in the y direction; u represents the horizontal frequency component in the frequency domain; v represents the vertical frequency component in the frequency domain; x and y represent the pixel coordinates in the spatial domain; Power spectral density E: describes the energy distribution of different frequency components in the image; E(u,v)=|F(u,v)| 2 Among them, E(u,v) represents the power of each frequency component in the frequency domain; |F(u,v)| represents the amplitude of each frequency component in the frequency domain; The high-frequency and low-frequency components of the image are extracted through a bandpass filter; High frequency component: satisfies u≥u c , v ≥ v c ,u c 1 / 2 of the horizontal frequency range; v c 1 / 2 of the vertical frequency range; High frequency component energy E high : Total energy E total : High frequency component energy ratio HF: The high-frequency component and the low-frequency component of the image reflect the edge and texture features of the image and the overall smoothness of the image; Autocorrelation feature: The autocorrelation function is used to describe the correlation between image pixels; Where K represents the total number of pixels in the image; R(Δx, Δy) represents the autocorrelation value between pixels at the displacement (Δx, Δy); f(x+Δx, y+Δy) represents the image after displacement; Histogram features: Histogram mean H Mean : Among them, H Mean represents the mean of the histogram image; L represents the number of gray levels; p(t) represents the probability of gray level t; Histogram variance H var : Among them, H var represents the histogram variance; Perform feature training on the intelligent denoising model.

4. The remote sensing data intelligent noise reduction and enhancement method based on deep learning according to claim 3 is characterized by: In step S2, the noise category C k Including Gaussian noise, impulse noise, Poisson noise and shot noise, the specific judgment method is: Select kurtosis Kur as a parameter in the statistical characteristics: Where Kur represents kurtosis; Indicates noise type C k The average value of the lower kurtosis; Indicates noise type C k The standard deviation under Select the energy ratio of high-frequency components in the frequency domain features: Wherein, HF represents the high frequency component energy ratio; Indicates noise type C k The average value of the energy ratio of the lower high frequency components; Indicates noise type C k The standard deviation of the energy ratio of the lower high frequency components; Select the zero-delay autocorrelation value as a parameter in the autocorrelation feature: set the values ​​of Δx and Δy to 0 to obtain the zero-delay autocorrelation value; Where AC represents the zero-delay autocorrelation value; Indicates noise type C k The average of the zero-delay autocorrelation values ​​under ; Indicates noise type C k The standard deviation of the zero-lag autocorrelation values ​​under ; In the histogram feature, the histogram mean is selected as the parameter: Among them, H var represents the histogram variance; Indicates noise type C k The average value of the histogram variance under ; Indicates noise type C k The standard deviation of the histogram variance under ; Calculate the probability: Among them, P(Kur), P(HF), P(AC) and P(H var ) represents the marginal probability of the feature; P(C k ) indicates noise type C k The prior probability of not observing the noise feature; P(Kur|C k ) indicates that in noise type C k The probability density function of kurtosis is: P(HF|C k ) indicates that in noise type C k The probability density function of the high-frequency component energy ratio is: P(AC|C k ) indicates that in noise type C k The probability density function of the zero-delay autocorrelation value under var |C k ) indicates that in noise type C k Below, the probability density function of the histogram mean; in, Indicates the post-verification noise type, including Gaussian noise, impulse noise, Poisson noise, and shot noise; By calculating the posterior probability of each noise type, and selecting the noise type with the largest posterior probability.

5. The remote sensing data intelligent noise reduction and enhancement method based on deep learning according to claim 4 is characterized in that: In step S2, the Gaussian noise is denoised using wavelet transform, and the specific steps are as follows: S5-1, decompose the image into different scales and directions through wavelet transform, and perform threshold processing on the wavelet coefficients; S5-2, performing wavelet transform on the image to obtain wavelet coefficients; Perform soft thresholding and hard thresholding on wavelet coefficients; S5-3. Perform inverse wavelet transform on the processed wavelet coefficients to reconstruct the image.

6. The remote sensing data intelligent noise reduction and enhancement method based on deep learning according to claim 4 is characterized in that: In step S3, the impulse noise is denoised using adaptive median filtering, the specific steps are: S6-1, initialize window size; S6-2, calculating the median, maximum and minimum values ​​of the pixels in the window; S6-3, adjusting the window size according to whether the pixel value is between the median and the extreme value; S6-4. Replace the central pixel with the median value of the adaptive window.

7. The remote sensing data intelligent noise reduction and enhancement method based on deep learning according to claim 4 is characterized in that: In step S3, the method for processing Poisson noise is to use Anscombe transform to transform the image containing Poisson noise into a Gaussian noise-like image, then use wavelet transform to reduce noise, and then perform inverse transform.

8. The remote sensing data intelligent noise reduction and enhancement method based on deep learning according to claim 1, characterized in that: In step S3, the image registration and merging adopts weighted average merging, specifically: in, Represents the pixel value of the merged image at (x, y); I a (x,y),I b (x,y) and I c (x, y) represents the pixel value of the visible light image, near infrared image and short wave infrared image at (x, y) respectively; w1, w2 and w3 represent the weights of the visible light image, near infrared image and short wave infrared image respectively.

9. The remote sensing data intelligent noise reduction and enhancement method based on deep learning according to claim 1, characterized in that: In step S4, the color correction and image enhancement are specifically as follows: The color correction is used to adjust the gain of each channel; select the white area in the image to perform white balance correction; use the reference image to perform color mapping so that the color distribution of the target image is consistent with the reference image; The image enhancement is used to enhance the contrast of the image; the edge and details of the image are enhanced through unsharp mask processing, and then converted to the HSV color space to adjust the color saturation and hue.

10. A remote sensing data intelligent denoising and enhancement system based on deep learning, which executes the remote sensing data intelligent denoising and enhancement method based on deep learning as claimed in claim 1, characterized in that: The system includes a data acquisition module, an intelligent noise reduction model module, an image registration and merging module, an image post-processing module and a result display module; The data acquisition module is used to collect remote sensing image data of the target area through satellite and UAV sensors; the intelligent noise reduction model module is used to process the collected remote sensing images, remove various types of noise, including Gaussian noise, impulse noise, Poisson noise and shot noise, and enhance the image quality; the image registration and merging module is used to register and merge images of multi-spectral bands to generate a comprehensive image containing more information and details; the image post-processing module is used to further optimize the image quality, enhance the details and contrast of the image, and perform color correction and image enhancement; the result display module is used to display the processed image to the user, provide visualization tools and interfaces, and support the user to further analyze and operate the image; The output end of the data acquisition module is connected to the input end of the intelligent denoising model module; the output end of the intelligent denoising model module is connected to the input end of the image registration and merging module; the output end of the image registration and merging module is connected to the input end of the image post-processing module; the output end of the image post-processing module is connected to the input end of the result display module.

Images