Go-pos scene adaptive remote sensing image super-resolution reconstruction system and method

Abstract

The invention relates to a Go-pos scene self-adaptive remote sensing image super-resolution reconstruction system and method, relates to the technical field of space optics, and solves the technical problems that when a traditional algorithm faces a remote sensing image with a complex landform, super-resolution reconstruction fails, the image edge reconstruction effect is poor, initial frame construction is incomplete, and subsequent iteration cannot be supported. The system comprises an input image loading module, a scene recognition module, a degradation model configuration module, an initial reconstruction NEDI module, a POCS iteration reconstruction module, an edge enhancement module and an output and storage module. According to the Go-pos scene adaptive remote sensing image super-resolution reconstruction system and method provided by the invention, a set of remote sensing image super-resolution reconstruction scheme with strong robustness, wide adaptability and excellent detail reconstruction effect is constructed through combination of scene recognition driven adaptive degradation modeling and an edge-oriented reconstruction strategy; and particularly, the method shows higher universality and reconstruction quality in diversified landform scenes.

Description

Technical Field

[0001] This invention relates to the field of space optics technology, and in particular to a Geo-pocs scene adaptive remote sensing image super-resolution reconstruction system and method. Background Technology

[0002] Super-resolution reconstruction of remote sensing images is a technique that utilizes mathematical and physical knowledge to process low-resolution images and generate high-resolution images using algorithms and neural networks. In research projects, remote sensing images directly captured by optical cameras often have low spatial resolution, making them insufficient for practical needs. In such cases, image processing techniques, such as super-resolution reconstruction, are required to improve image quality and increase spatial resolution to meet project requirements. Since its inception, this method has attracted a large number of researchers. The methods in various application areas are complex and diverse. Especially in recent years, the integration with deep learning has led to an explosive growth in super-resolution reconstruction publications. However, along with the expansion of its application areas, problems have also arisen.

[0003] While traditional model-driven reconstruction methods have achieved some success in reconstruction, they often rely too heavily on prior image information, resulting in poor edge reconstruction. Furthermore, they lack universality when dealing with diverse terrains or complex landforms, and the complexity of influencing factors can lead to super-resolution reconstruction failure. Moreover, the need for optimal parameter tuning for different images and devices severely limits the performance and applicability of these methods. Deep learning methods offer better metrics, but they are essentially "black box" processes with poor interpretability. The complexity of neural networks and their massive computational demands, high GPU usage, and large-scale training detached from real-world models are difficult to meet application requirements. Currently, most mature algorithms are based on spatial convex set projection (POCS), maximum a posteriori probability (MAP), and iterative back projection. Among these, convex set projection has been extensively studied due to its intuitive algorithm principle, flexible imaging model, and ease of incorporating prior information. Its main drawbacks are large derivatives of edge grayscale values, incomplete PSF correction, poor initial estimation of high-resolution frames, and lack of universality for different terrain types. Summary of the Invention

[0004] This invention aims to address the technical problems of traditional algorithms failing in super-resolution reconstruction of remote sensing images with complex terrain, exhibiting poor image edge reconstruction results, and failing to support subsequent iterations due to incomplete initial frame construction. It provides a Geo-pocs scene adaptive remote sensing image super-resolution reconstruction system and method.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] A Geo-POCS scene-adaptive remote sensing image super-resolution reconstruction system includes, in sequence: an input image loading module, a scene recognition module, a degradation model configuration module, an initial reconstruction NEDI module, a POCS iterative reconstruction module, an edge enhancement module, and an output and storage module; wherein:

[0007] The input image loading module is used to preprocess an input remote sensing image file, determine whether it is a color image; if it is a color image, it is converted to grayscale and then converted to double type for normalization;

[0008] The scene recognition module is used to infer the type of terrain scene from the content of remote sensing images by using the gray-scale statistical features, edge structure features, and texture complexity features of the images.

[0009] The degradation model configuration module is used to call the corresponding degradation model parameters according to the scene to simulate low-resolution images;

[0010] The initial reconstruction NEDI module is used to perform interpolation operations on the low-resolution image by upsampling using the NEDI method, so as to obtain a high-resolution initial reference frame with the target resolution.

[0011] The POCS iterative reconstruction module is used to process each pixel in the low-resolution image one by one using the motion parameter estimation process. It finds the center coordinates of the corresponding PSF window of the corresponding pixel in the high-resolution initial reference frame through block matching. It then performs point spread function correction based on the residual pixel values ​​between the simulated and real low-resolution images.

[0012] The edge enhancement module is used to reconstruct edge details. It uses the Canny operator to extract the image edge prior and the Sobel operator to extract the image edge gradient before performing edge-guided enhancement to preserve structural details and improve perceptual quality.

[0013] The output and save module is used to iteratively complete the output of the final super-resolution image and save it.

[0014] In the above technical solution, the scene recognition module is specifically used for:

[0015] Assuming the image is I(x,y), the image's classification is determined by solving for the following eigenvalues: the mean gray value μ, which characterizes the overall brightness of the image.

[0016]

[0017] in:

[0018] H: Image height;

[0019] W: Width of the image;

[0020] x: row index of the pixel;

[0021] y: Column index of the pixel;

[0022] I(x,y): The gray value at position (x,y);

[0023] The grayscale standard deviation σ, which characterizes the degree of drastic change in image brightness:

[0024]

[0025] Information entropy H represents the complexity of an image:

[0026]

[0027] in:

[0028] μ: Mean gray level;

[0029] i: Index of grayscale level;

[0030] p i The probability that a pixel with a gray value of i will appear in the entire image;

[0031] and edge density E d :

[0032]

[0033] In the above technical solution, the degradation model configuration module is specifically used to obtain degradation models for different landforms, wherein:

[0034] Urban topography: Subpixel translation is used for motion transformation; the blur kernel is a convolution of a motion blur kernel and a Gaussian kernel; and additive Gaussian noise is used for the noise model.

[0035] Desert landforms: The motion transformation adopts an approximate identity transformation, and the fuzzy kernel adopts a noise model with low-intensity Gaussian noise;

[0036] Mountainous terrain: Affine transformation is used for motion transformation, and low-intensity Gaussian noise is used for the fuzzy kernel; the noise model adopts a Gaussian mixture + Poisson model.

[0037] Forest topography: The motion transformation adopts non-rigid perturbation with an elliptical Gaussian fuzzy kernel; the noise model adopts multiplicative noise.

[0038] Water body topography: The motion transformation adopts local perturbation displacement, and the fuzzy kernel adopts slight motion + Gaussian; the noise model adopts Poisson noise.

[0039] In the above technical solution, the initial reconstruction of the NEDI module is specifically used for:

[0040] Selective activation for edge-sensitive scenes: NEDI is used if the scene is a city or mountain; Bicubic interpolation is used if it is a body of water or desert to prevent false edges. The interpolation results are used for multiple rounds of POCS reprojection correction to ensure the quality of the initial structure estimation and provide a basis for subsequent edge enhancement and residual backprojection. It also provides a high-quality super-resolution initial estimation image, providing a good starting point for subsequent POCS iterations.

[0041] In the above technical solution, the POCS iterative reconstruction module is specifically used for:

[0042] Initial image construction uses the initial estimated image obtained from NEDI or Bicubic as the reference frame; the low-resolution image I, which has been classified according to terrain features, is used as the reference frame. LR The image is fed into the corresponding degradation model to generate a simulated low-resolution image. Then, the error residual R between the simulated low-resolution image and the real low-resolution image was calculated. (k) :

[0043]

[0044] The error residual R (k) Upsample back to the original high-resolution image, correct the high-resolution initial frame, and use the backprojection residuals to update the high-resolution map;

[0045] The nonnegativity constraint in the spatial domain prior requires that all pixels be greater than zero during projection:

[0046]

[0047] in:

[0048] Set of nonnegativity constraints;

[0049] x: represents an image;

[0050] x(i,j): The pixel value of image x at the i-th row and j-th column;

[0051] This holds true for all row indexes i and column indexes j;

[0052] Amplitude constraints can prevent abnormal or overflowing pixel values ​​during reconstruction. Pixel values ​​should be within the range of [0, 255].

[0053]

[0054] in:

[0055] Amplitude constraint set;

[0056] The definition of energy constraint is:

[0057]

[0058] in:

[0059] C E Energy constraint set;

[0060] X: Represents an image signal or matrix;

[0061] The squared L2 norm of an image, which is the sum of the squares of all pixels;

[0062] X m,n : The pixel value of the image in the m-th row and n-th column;

[0063] E max : The maximum allowable energy threshold;

[0064] The total variation constraint is defined as follows:

[0065]

[0066] in:

[0067] TV(x): Total variation of image x, used to measure the sum of gradients of the image;

[0068] x i,j : The value of a pixel in the i-th row and j-th column;

[0069] x i+1,j -x i,j The difference between pixels in the vertical direction;

[0070] x i,j+1 -x i,j The difference between pixels in the horizontal direction;

[0071] The current estimated image is mapped to a set that satisfies various priors for projection correction, and then iterated, where k represents the reconstructed image after the k-th iteration;

[0072] The final result is a high-resolution remote sensing image with clear edges, high realism, and excellent reconstruction effect.

[0073] In the above technical solution, the edge enhancement module is specifically used for:

[0074] The binary edge map E is extracted using Canny, and the image gradient is calculated.

[0075] G x =Sobel x (X (k) ),G y =Sobel y (X(k) )

[0076]

[0077] in:

[0078] X(k): The image obtained in the k-th iteration;

[0079] Sobel x : Perform convolution using the Sobel operator in the horizontal direction to extract the gradient in the horizontal direction;

[0080] Sobel y : Perform convolution using the Sobel operator in the vertical direction to extract the gradient in the vertical direction;

[0081] G x : The gradient component of the image in the x-direction;

[0082] G y : The gradient component of the image in the y-direction;

[0083] G: Overall gradient magnitude map of the image, reflecting edge intensity;

[0084] Perform gradient normalization:

[0085]

[0086] in:

[0087] ε: A very small positive number used to avoid the denominator being 0, thus ensuring numerical stability;

[0088] max(G): Normalizes all gradient values ​​to the range [0,1] by the maximum value;

[0089] G norm : Normalized gradient plot;

[0090] Finally, gradient edge guidance enhancement is performed:

[0091] X (k+1) =X (k) +γ·E·G norm ·255

[0092] in:

[0093] X (k+1) This is the image from the (k+1)th iteration;

[0094] X (k) The image for the current iteration;

[0095] E is the edge map extracted by the Canny operator, ensuring that enhancement is only applied to the true edge region;

[0096] γ is the scene adaptive enhancement coefficient;

[0097] In each POCS iteration, edge enhancement is inserted into the main process as a projection operation, and pixel value range is limited.

[0098] A Geo-pocs scene-adaptive remote sensing image super-resolution reconstruction method applicable to the aforementioned Geo-pocs scene-adaptive remote sensing image super-resolution reconstruction system includes the following steps:

[0099] Step 1: Read a frame of remote sensing image, preprocess it into a grayscale image and normalize it;

[0100] Perform image preprocessing by reading the input remote sensing image and normalizing the image grayscale values ​​to a fixed range;

[0101] Step 2: Extract image features and use a rule tree to classify the terrain scene in the image;

[0102] Automatic scene recognition is performed, statistical features of images are extracted, and rule trees are used to classify images into scenes and identify their respective landform types;

[0103] Step 3: Based on the scene, call the corresponding degradation model parameters to simulate a low-resolution image;

[0104] Based on different terrains, specific downsampling methods are selected, and after obtaining the noise level and edge enhancement weight, the image is sent to its corresponding degradation model to simulate a low-resolution image.

[0105] Step 4: Perform initial upsampling using NEDI interpolation;

[0106] The NEDI method is used to upsample the low-resolution image to complete the interpolation operation, resulting in a high-resolution initial reference frame with the target resolution.

[0107] Step 5: Iteratively execute the convex set projection reconstruction process;

[0108] The motion parameter estimation process is used to process each pixel in the low-resolution image one by one. Block matching is used to find the center coordinates of the corresponding PSF window of the corresponding pixel in the high-resolution initial frame. Point spread function correction is performed based on the residual of pixel values ​​between the simulated and real low-resolution images.

[0109] Step 6: Extract edge priors in each iteration and perform gradient-guided enhancement;

[0110] Edge detail reconstruction is performed by using the Canny operator to extract image edge priors and Sobel to extract image edge gradients, followed by edge-guided enhancement to preserve structural details and improve perceptual quality.

[0111] Step 7: Iterate through the process, output the final super-resolution image, and save it;

[0112] Iterative corrections are performed based on convex set projection theory, and image details are reconstructed through back projection error compensation, resulting in a high-resolution image with enhanced details and clear edges.

[0113] The present invention has the following beneficial effects:

[0114] The Geo-pocs scene-adaptive remote sensing image super-resolution reconstruction system and method of the present invention solves the failure problems that often occur in traditional super-resolution reconstruction algorithms when dealing with remote sensing images of complex terrains, such as poor image edge reconstruction, incomplete initial high-resolution frame construction, and inability to support subsequent iteration processes. The method and system of the present invention have the following significant advantages:

[0115] First, an automatic scene recognition algorithm is introduced. By extracting features from remote sensing images, it automatically determines the terrain type of the image, including typical natural and man-made environments such as deserts, cities, water bodies, mountains, and forests. This module greatly enhances the algorithm's versatility and intelligence, enabling the reconstruction process to dynamically adjust strategies based on scene characteristics.

[0116] Secondly, based on the identification results, targeted downsampling methods, fuzzy kernel functions, and noise modeling parameters were designed for different terrains to more realistically simulate the sensor imaging degradation process. This scene-adaptive configuration mechanism effectively improves the realism of the degradation model and the reconstruction accuracy.

[0117] In constructing the initial high-resolution frame, the NEDI edge-guided interpolation algorithm is adopted, which can effectively preserve the edge structure of the original image, avoid introducing false textures and interpolation noise, and provide a high-quality prior foundation for subsequent reconstruction.

[0118] Subsequently, motion estimation and point spread function (PSF) correction are performed using the initial frame to further accurately align the spatial relationships between multiple frames and reduce the interference of motion artifacts on the reconstruction results.

[0119] Furthermore, an edge prior guidance mechanism was introduced during the reconstruction process. Clear edge regions were extracted using the Canny edge detection algorithm, and then a gradient enhancement strategy was employed to strengthen the expression of edge information during the iterative process, achieving an image reconstruction optimization similar to an edge attention mechanism. This design significantly improved the representation of edge details and structure in the reconstructed image.

[0120] In summary, the Geo-pocs scene-adaptive remote sensing image super-resolution reconstruction system and method of the present invention combines scene recognition-driven adaptive degradation modeling with an edge-oriented reconstruction strategy to construct a robust, adaptable, and detail-oriented remote sensing image super-resolution reconstruction scheme, which exhibits stronger versatility and reconstruction quality, especially in diverse terrain scenes. Attached Figure Description

[0121] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0122] Figure 1 This is a schematic diagram of the Geo-pocs scene adaptive remote sensing image super-resolution reconstruction system of the present invention.

[0123] Figure 2 This is a schematic diagram illustrating the steps of the Geo-pocs scene adaptive remote sensing image super-resolution reconstruction method of the present invention.

[0124] Figure 3 This is a flowchart illustrating the steps of the iterative execution of convex set projection (POCS) reconstruction process in the Geo-pocs scene adaptive remote sensing image super-resolution reconstruction method of the present invention. Detailed Implementation

[0125] The inventive concept of this invention is as follows:

[0126] The Geo-pocs scene-adaptive remote sensing image super-resolution reconstruction system and method of this invention determines the terrain type of a remote sensing image based on its features: desert, city, water body, mountain, or forest. It analyzes terrain features to design different downsampling methods, blur kernels, and noise levels. Based on the classification results, it dynamically loads and performs scene-adaptive configuration. The reconstruction module constructs an initial reference frame using NEDI interpolation to ensure good image edges without introducing other noise. The initial reference frame undergoes motion parameter estimation and point spread function correction. The Canny edge detection algorithm is used to extract edge priors, followed by gradient-guided enhancement to achieve edge attention mechanism reconstruction. This results in remote sensing image super-resolution reconstruction with strong terrain universality and good reconstruction effect.

[0127] The present invention will now be described in detail with reference to the accompanying drawings.

[0128] This invention proposes a Geo-pocs scene-adaptive remote sensing image super-resolution reconstruction system and method. By combining traditional image interpolation methods with convex set projection algorithms, and introducing a terrain scene classification mechanism and edge-guided enhancement strategy, it achieves refined image reconstruction for different terrain types. This method does not require deep learning neural network training, possesses advantages such as high interpretability and low computational resource requirements, and is particularly suitable for resource-constrained remote sensing image processing scenarios.

[0129] like Figure 1 As shown, the Geo-pocs scene adaptive remote sensing image super-resolution reconstruction system of the present invention includes the following components connected in sequence:

[0130] S1: Input image loading module;

[0131] The input image loading module preprocesses an input remote sensing image file, determining whether it is a color image. If it is a color image, it is converted to grayscale and then converted to double type for normalization. The image grayscale values ​​are normalized to the range [0, 255] for consistency in subsequent processing.

[0132] S2: Scene recognition module;

[0133] The scene recognition module leverages the grayscale statistical features, edge structure features, and texture complexity features of remote sensing images to construct a rule-based image classification mechanism, inferring the terrain scene category from the image content. It calculates the image's information entropy (representing texture complexity), edge density, and the mean and standard deviation of grayscale values ​​to automatically identify image scene types, such as cities, mountains, forests, water bodies, or deserts. It requires no training set or deep network, using only statistical feature rule trees for classification. Compared to remote sensing semantic segmentation methods based on convolutional neural networks, this invention offers advantages such as no need for a training set, low computational complexity, and high interpretability, making it suitable for embedding into traditional algorithm workflows.

[0134] Assuming the image is I(x,y), the image's classification is determined by solving for the following eigenvalues: the mean gray value μ, which characterizes the overall brightness of the image.

[0135]

[0136] in

[0137] H: Image height (in pixels);

[0138] W: Image width (in pixels);

[0139] x: row index of the pixel;

[0140] y: Column index of the pixel;

[0141] I(x,y): Gray value (pixel brightness) at position (x,y);

[0142] The grayscale standard deviation σ, which characterizes the degree of drastic change in image brightness:

[0143]

[0144] Information entropy H represents the complexity of an image:

[0145]

[0146] in:

[0147] μ: Average gray level (average brightness);

[0148] i: Index of grayscale level;

[0149] p i The probability that a pixel with a gray value of i will appear in the entire image;

[0150] and edge density E d :

[0151]

[0152] Edge maps are obtained through Canny operator detection, edge density is calculated by counting the percentage of white pixels, and pixel probabilities p are estimated using histograms. i Information entropy is obtained, and after acquiring all feature values, they are compared based on terrain features. Urban areas have dense buildings, complex structures, and dense boundaries, resulting in the highest entropy value. Mountain ranges have varying elevations, causing large differences in illumination, resulting in rough textures but non-uniform edges and moderate edge density. Forest areas are generally dark with low mean grayscale. Water bodies have extremely low edge density due to their low-contrast areas and small standard deviation of grayscale. Deserts have low edge density, but their entropy is slightly higher than that of water bodies. They also have high albedo, high brightness, and relatively high mean grayscale. This establishes a rule-based classification mechanism that can infer the terrain scene category from the image content for classification.

[0153] S3: Degradation model configuration module;

[0154] The degradation model configuration module is used to call the corresponding degradation model parameters according to the scene to simulate low-resolution images;

[0155] In the actual imaging process of remote sensing images, traditional degradation models show that different landform scenes (such as cities, water bodies, forests, etc.) are affected by factors such as imaging system parameters, terrain structure, surface materials, and climate conditions during the shooting process, resulting in significant differences in their image degradation patterns.

[0156] Super-resolution methods typically assume a single, simplified degradation model (such as isotropic Gaussian blur + added Gaussian noise + fixed downsampling).

[0157] For any scene c∈C, its low-resolution image degradation process is modeled as follows:

[0158]

[0159] in:

[0160] Y c Degradation result;

[0161] X: Original high-resolution image;

[0162] Geometric perturbations / motion transformations (such as translation, affine transformation, non-rigid body transformation);

[0163] Spatial blurring operations lead to degradation of the modeling optical system;

[0164] Downsampling operation;

[0165] N c Imaging noise.

[0166] To obtain degradation models specific to different landforms, the principles of several scenarios are analyzed:

[0167] The terrain features in remote sensing images are divided into five categories: cities, water bodies, mountains, forests, and deserts. The underlying principles of each scene are analyzed, leading to the following conclusions:

[0168] The city has a dense regional structure with clear edges, and there are slight camera movements and imaging system defocusing caused by building obstruction, vehicles, and wind speed. Motion transformation considers sub-pixel translation. The blur kernel uses a convolution of a motion blur kernel and a Gaussian kernel, and the noise model considers additive Gaussian noise.

[0169] The desert region has sparse texture, uniform imaging illumination, and small parallax variations, but thermal air disturbances and lens distortion cause slight optical defocusing blur. Motion transformations are considered to be approximately identical transformations, and the blur kernel uses a noise model with low-intensity Gaussian noise.

[0170] The mountainous terrain is highly variable, resulting in strong shadows, uneven lighting, and affine distortion and blurring caused by tilted shooting, along with high imaging noise. Affine transformation is considered for motion transformation, and the blur kernel formula is the same as for the desert region. A Gaussian mixture plus Poisson model is used for noise modeling.

[0171] Forest areas possess fine, natural textures. The non-rigid local deformation caused by tree canopy movement leads to spatial correlation ambiguity, often accompanied by multiplicative noise. Motion transformations consider non-rigid perturbations and employ an elliptical Gaussian blur kernel. The noise model uses multiplicative noise (speckle noise).

[0172] The water surface exhibits wave-like behavior, and specular reflection and disturbance cause local sub-pixel displacement. Simultaneously, the highlight areas in the image are dominated by Poisson photon noise. The motion transformation is a local perturbation displacement (small-amplitude sliding), and the blur kernel uses a slight motion plus Gaussian approach. Poisson noise is selected as the noise model.

[0173] Using the scene recognition module, images classified according to terrain are matched with the corresponding degradation model.

[0174] S4: Initial reconstruction of the NEDI module;

[0175] The initial reconstruction NEDI module is used to perform interpolation operations on low-resolution image sampling using the New Edge-Directed Interpolation (NEDI) method to obtain a high-resolution initial reference frame (HR, High Resolution) with the target resolution.

[0176] NEDI (New Edge-Directed Interpolation) is an interpolation method based on locally adaptive edge direction information, which can effectively enhance the quality of edge reconstruction without introducing external training priors. It selects a subset of known pixels from a low-resolution image as training samples; calculates the covariance matrix (structural information) within this region; assumes that the same edge structure is maintained in the high-resolution image; and estimates unknown pixels using the minimum mean square error (LMMSE).

[0177] In this invention, NEDI is used in the initial image construction module, and adaptation optimizations are made in the following aspects:

[0178] Selective activation for edge-sensitive scenes: NEDI is used if the scene is a city or mountain; Bicubic interpolation is used if it is a body of water or desert to prevent false edges. The interpolation results are used for multiple rounds of POCS reprojection correction to ensure the quality of the initial structure estimation and provide a basis for subsequent edge enhancement and residual backprojection. It also provides a high-quality super-resolution initial estimation image, providing a good starting point for subsequent POCS iterations.

[0179] S5: POCS iterative reconstruction module;

[0180] The POCS iterative reconstruction module is used to process each pixel in the low-resolution image one by one using the motion parameter estimation process. It finds the center coordinates of the corresponding PSF window of the corresponding pixel in the high-resolution initial frame through block matching. It then performs point spread function correction based on the residual pixel values ​​between the simulated and real low-resolution images.

[0181] The convex set projection method models a set of images satisfying different observation conditions or prior knowledge as a set of convex sets, where the target image should fall within the intersection of these sets. Super-resolution reconstruction is achieved by continuously approximating the intersection of multiple convex sets (such as data consistency constraints, prior structure constraints, etc.). Each iteration mainly completes two core operations: data consistency projection (data constraints) and prior structure-guided projection, as detailed in the following process: Figure 3 As shown, it specifically includes:

[0182] First, initial image construction is performed, using the initial estimated image obtained from NEDI or Bicubic as the reference frame. The low-resolution image I, after being classified according to terrain features, is then used. LR The image is fed into the corresponding degradation model to generate a simulated low-resolution image. Then, the error residual R between the simulated low-resolution image and the real low-resolution image was calculated. (k) :

[0183]

[0184] The error residual R (k) The original high-resolution image is upsampled back to correct the initial high-resolution frame, and the backprojection residuals are used to update the high-resolution image. The point spread function correction process requires the use of convex set constraints for each pixel, requiring that the absolute value of the residuals be limited within a pre-set boundary. Artificial priors are introduced, such as magnitude constraints, energy constraints, and model similarity constraints.

[0185] The nonnegativity constraint in the spatial domain prior requires that all pixels be greater than zero during projection:

[0186]

[0187] in:

[0188] Set of nonnegativity constraints;

[0189] x: represents an image;

[0190] x(i,j): The pixel value of image x at the i-th row and j-th column;

[0191] This holds true for all row indexes i and column indexes j;

[0192] Amplitude constraints can prevent abnormal or overflowing pixel values ​​during reconstruction. Pixel values ​​should be within the range of [0, 255].

[0193]

[0194] in:

[0195] Amplitude constraint set;

[0196] Energy constraints require that the reconstructed image cannot contain excessively high energy, which can suppress high-frequency oscillations. The constraint is defined as follows:

[0197]

[0198] in:

[0199] C E Energy constraint set;

[0200] X: Represents an image signal or matrix;

[0201] The squared L2 norm of an image, which is the sum of the squares of all pixels (energy);

[0202] X m,n : The pixel value of the image in the m-th row and n-th column;

[0203] E max : The maximum allowable energy threshold;

[0204] Total Variation Projection can suppress artifacts and noise in reconstructed images, preserve structural boundaries, suppress drastic changes in the image, maintain smooth regions, and preserve edges.

[0205]

[0206] in:

[0207] TV(x): Total variation of image x, used to measure the sum of gradients of the image;

[0208] x i,j : The value of a pixel in the i-th row and j-th column;

[0209] x i+1,j -x i,j The difference between pixels in the vertical direction (row direction gradient);

[0210] x i,j+1 -x i,j The difference between pixels in the horizontal direction (column direction gradient);

[0211] The current estimated image is mapped to a set that satisfies various priors for projection correction, and then iterated, where k represents the reconstructed image after the k-th iteration.

[0212] The final result is a high-resolution remote sensing image with clear edges, high realism, and excellent reconstruction effect.

[0213] S6: Edge enhancement module;

[0214] The edge enhancement module is used to reconstruct edge details. It uses the Canny operator to extract the image edge prior and the Sobel operator to extract the image edge gradient before performing edge-guided enhancement to preserve structural details and improve perceptual quality.

[0215] Traditional interpolation methods are prone to causing edge blurring and texture loss during reconstruction; while in iterative reconstruction methods, without structural guidance, edges may become too smooth or falsely enhanced.

[0216] The edge enhancement module performs edge-guided enhancement, utilizing the prior edge structure of the image to dynamically constrain and guide image detail enhancement in each iteration, thereby improving edge fidelity and overall structural clarity. Edge-guided enhancement is a process of detail compensation based on spatial prior information. Essentially, it involves: constructing a spatial weight map (i.e., an edge map) to characterize important structural regions in the image; using this edge map as a weighting factor to adjust the gradient enhancement intensity; and applying enhancement only to structurally significant regions to prevent the amplification of false textures and noise.

[0217] The specific steps are as follows:

[0218] The binary edge map E is extracted using Canny, and the image gradient is calculated.

[0219] G x =Sobel x (X (k) ),G y =Sobel y (X (k) )

[0220]

[0221] in:

[0222] X(k): The image obtained in the k-th iteration;

[0223] Sobel x : Perform convolution using the Sobel operator in the horizontal direction to extract the gradient in the horizontal direction;

[0224] Sobel y : Perform convolution using the Sobel operator in the vertical direction to extract the gradient in the vertical direction;

[0225] G x The gradient component of the image in the x-direction (horizontal gradient);

[0226] G y The gradient component of the image in the y-direction (vertical gradient);

[0227] G: Overall gradient magnitude map of the image, reflecting edge intensity;

[0228] Perform gradient normalization:

[0229]

[0230] in:

[0231] ε: A very small positive number used to avoid the denominator being 0, thus ensuring numerical stability;

[0232] max(G): Normalizes all gradient values ​​to the range [0,1] by the maximum value;

[0233] G norm : Normalized gradient plot;

[0234] Finally, gradient edge guidance enhancement is performed:

[0235] X (k+1) =X (k) +γ·E·G norm ·255

[0236] in:

[0237] X (k+1) This is the image from the (k+1)th iteration;

[0238] X (k) The image for the current iteration;

[0239] E is the edge map extracted by the Canny operator, ensuring that enhancement is only applied to the true edge region;

[0240] γ is the scene adaptive enhancement coefficient;

[0241] In each POCS iteration, edge enhancement is inserted into the main process as a projection operation, and pixel value range is limited.

[0242] S7: Output and save module;

[0243] The output and save module is used to iterate and complete the final super-resolution image output and save it.

[0244] like Figure 2 As shown, the Geo-pocs scene adaptive remote sensing image super-resolution reconstruction method of the present invention includes the following steps:

[0245] Step 1: Read a frame of remote sensing image, preprocess it into a grayscale image and normalize it;

[0246] Image preprocessing is performed by reading the input remote sensing image and normalizing the image grayscale values ​​to a fixed range.

[0247] Step 2: Extract image features (information entropy, edge density, grayscale statistics) and use a rule tree to classify the terrain scene of the image;

[0248] Automatic scene recognition is performed, and statistical features of the image (including information entropy, edge density, gray mean and standard deviation, etc.) are extracted. The image is then classified into scenes using a rule tree to identify the type of terrain to which it belongs.

[0249] Step 3: Based on the scene, call the corresponding degradation model parameters to simulate a low-resolution image;

[0250] Specific downsampling methods are selected based on different terrains (e.g., Bicubic is used for cities, and the blur kernel considers vehicle / wind motion blur + slight defocus blur). After obtaining the noise level and edge enhancement weight, the image is sent to its corresponding degradation model to simulate low-resolution images (LR).

[0251] Step 4: Perform initial upsampling using NEDI interpolation;

[0252] The NEDI (New Edge-Directed Interpolation) method is used to upsample the low-resolution image to perform interpolation, resulting in a high-resolution initial reference frame (HR) with the target resolution.

[0253] Step 5: Iteratively execute the convex set projection (POCS) reconstruction process;

[0254] The motion parameter estimation process processes each pixel in the low-resolution image one by one, finding the center coordinates of the corresponding PSF window in the high-resolution initial frame through block matching. Point spread function correction is then performed based on the residual pixel values ​​between the simulated and real low-resolution images.

[0255] Step 6: Extract edge priors in each iteration and perform gradient-guided enhancement;

[0256] Edge detail reconstruction is performed by using the Canny operator to extract image edge priors and Sobel to extract image edge gradients, followed by edge-guided enhancement to preserve structural details and improve perceptual quality.

[0257] Step 7: Iterate through the process, output the final super-resolution image, and save it;

[0258] Iterative corrections are performed based on convex set projection theory, and image details are reconstructed through back projection error compensation, resulting in a high-resolution image with enhanced details and clear edges.

[0259] The Geo-pocs scene-adaptive remote sensing image super-resolution reconstruction system and method of the present invention combines scene recognition-driven adaptive degradation modeling with an edge-oriented reconstruction strategy to construct a robust, adaptable, and detail-reconstruction-effective remote sensing image super-resolution reconstruction scheme, which exhibits stronger versatility and reconstruction quality, especially in diverse terrain scenes.

[0260] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A Geo-pocs scene-adaptive remote sensing image super-resolution reconstruction system, characterized in that, It includes, in sequence: an input image loading module, a scene recognition module, a degradation model configuration module, an initial reconstruction NEDI module, a POCS iterative reconstruction module, an edge enhancement module, and an output and saving module; wherein: The input image loading module is used to preprocess an input remote sensing image file, determine whether it is a color image; if it is a color image, it is converted to grayscale and then converted to double type for normalization; The scene recognition module is used to infer the type of terrain scene from the content of remote sensing images by using the gray-scale statistical features, edge structure features, and texture complexity features of the images. The degradation model configuration module is used to call the corresponding degradation model parameters according to the scene to simulate low-resolution images; The initial reconstruction NEDI module is used to perform interpolation operations on the low-resolution image by upsampling using the NEDI method, so as to obtain a high-resolution initial reference frame with the target resolution. The POCS iterative reconstruction module is used to process each pixel in the low-resolution image one by one using the motion parameter estimation process. It finds the center coordinates of the corresponding PSF window of the corresponding pixel in the high-resolution initial reference frame through block matching. It then performs point spread function correction based on the residual pixel values ​​between the simulated and real low-resolution images. The edge enhancement module is used to reconstruct edge details. It uses the Canny operator to extract the image edge prior and the Sobel operator to extract the image edge gradient before performing edge-guided enhancement to preserve structural details and improve perceptual quality. The output and save module is used to iteratively complete the output of the final super-resolution image and save it.

2. The Geo-pocs scene adaptive remote sensing image super-resolution reconstruction system according to claim 1, characterized in that, The scene recognition module is specifically used for: Assuming the image is I(x,y), the image's classification is determined by solving for the following eigenvalues: the mean gray value μ, which characterizes the overall brightness of the image. in: H: Image height; W: Width of the image; x: row index of the pixel; y: Column index of the pixel; I(x,y): The gray value at position (x,y); The grayscale standard deviation σ, which characterizes the degree of drastic change in image brightness: Information entropy H represents the complexity of an image: in: μ: Mean gray level; i: Index of grayscale level; p i The probability that a pixel with a gray value of i will appear in the entire image; and edge density E d :

3. The Geo-pocs scene adaptive remote sensing image super-resolution reconstruction system according to claim 1, characterized in that, The degradation model configuration module is specifically used to obtain degradation models for different landforms, including: Urban topography: Subpixel translation is used for motion transformation; the blur kernel is a convolution of a motion blur kernel and a Gaussian kernel; and additive Gaussian noise is used for the noise model. Desert landforms: The motion transformation adopts an approximate identity transformation, and the fuzzy kernel adopts a noise model with low-intensity Gaussian noise; Mountainous terrain: Affine transformation is used for motion transformation, and low-intensity Gaussian noise is used for the fuzzy kernel; the noise model adopts a Gaussian mixture + Poisson model. Forest topography: The motion transformation adopts non-rigid perturbation with an elliptical Gaussian fuzzy kernel; the noise model adopts multiplicative noise. Water body topography: The motion transformation adopts local perturbation displacement, and the fuzzy kernel adopts slight motion + Gaussian; the noise model adopts Poisson noise.

4. The Geo-pocs scene adaptive remote sensing image super-resolution reconstruction system according to claim 1, characterized in that, The initial reconstruction of the NEDI module is specifically used for: Selective activation for edge-sensitive scenes: NEDI is used if the scene is a city or mountain; Bicubic interpolation is used if it is a body of water or desert to prevent false edges. The interpolation results are used for multiple rounds of POCS reprojection correction to ensure the quality of the initial structure estimation and provide a basis for subsequent edge enhancement and residual backprojection. It also provides a high-quality super-resolution initial estimation image, providing a good starting point for subsequent POCS iterations.

5. The Geo-pocs scene adaptive remote sensing image super-resolution reconstruction system according to claim 1, characterized in that, The POCS iterative reconstruction module is specifically used for: Initial image construction uses the initial estimated image obtained from NEDI or Bicubic as the reference frame; the low-resolution image I, which has been classified according to terrain features, is used as the reference frame. LR The image is fed into the corresponding degradation model to generate a simulated low-resolution image. Then, the error residual R between the simulated low-resolution image and the real low-resolution image was calculated. (k) : The error residual R ( k) Upsample back to the original high-resolution image, correct the high-resolution initial frame, and use the backprojection residual to update the high-resolution map; The nonnegativity constraint in the spatial domain prior requires that all pixels be greater than zero during projection: in: Set of nonnegativity constraints; x: represents an image; x(i,j): The pixel value of image x at the i-th row and j-th column; This holds true for all row indexes i and column indexes j; Amplitude constraints can prevent abnormal or overflowing pixel values ​​during reconstruction. Pixel values ​​should be within the range of [0, 255]. in: Amplitude constraint set; The definition of energy constraint is: in: C E Energy constraint set; X: Represents an image signal or matrix; The squared L2 norm of an image, which is the sum of the squares of all pixels; X m,n : The pixel value of the image in the m-th row and n-th column; E max : The maximum allowable energy threshold; The total variation constraint is defined as follows: in: TV(x): Total variation of image x, used to measure the sum of gradients of the image; x i,j : The value of a pixel in the i-th row and j-th column; x i+1,j -x i,j The difference between pixels in the vertical direction; x i,j+1 -x i,j The difference between pixels in the horizontal direction; The current estimated image is mapped to a set that satisfies various priors for projection correction, and then iterated, where k represents the reconstructed image after the k-th iteration; The final result is a high-resolution remote sensing image with clear edges, high realism, and excellent reconstruction effect.

6. The Geo-pocs scene adaptive remote sensing image super-resolution reconstruction system according to claim 1, characterized in that, The edge enhancement module is specifically used for: The binary edge map E is extracted using Canny, and the image gradient is calculated. G x =Sobel x (X (k) ),G y =Sobel y (X (k) ) in: X(k): The image obtained in the k-th iteration; Sobel x : Perform convolution using the Sobel operator in the horizontal direction to extract the gradient in the horizontal direction; Sobel y : Perform convolution using the Sobel operator in the vertical direction to extract the gradient in the vertical direction; G x : The gradient component of the image in the x-direction; G y : The gradient component of the image in the y-direction; G: Overall gradient magnitude map of the image, reflecting edge intensity; Perform gradient normalization: in: ε: A very small positive number used to avoid the denominator being 0, thus ensuring numerical stability; max(G): Normalizes all gradient values ​​to the range [0,1] by the maximum value; G norm : Normalized gradient plot; Finally, gradient edge guidance enhancement is performed: X (k+1) =X (k) +γ·E·G norm ·255 in: X (k+1) This is the image from the (k+1)th iteration; X (k) The image for the current iteration; E is the edge map extracted by the Canny operator, ensuring that enhancement is only applied to the true edge region; γ is the scene adaptive enhancement coefficient; In each POCS iteration, edge enhancement is inserted into the main process as a projection operation, and pixel value range is limited.

7. A Geo-pocs scene adaptive remote sensing image super-resolution reconstruction method applicable to the Geo-pocs scene adaptive remote sensing image super-resolution reconstruction system of claim 1, characterized in that, Includes the following steps: Step 1: Read a frame of remote sensing image, preprocess it into a grayscale image and normalize it; Perform image preprocessing by reading the input remote sensing image and normalizing the image grayscale values ​​to a fixed range; Step 2: Extract image features and use a rule tree to classify the terrain scene in the image; Automatic scene recognition is performed, statistical features of images are extracted, and rule trees are used to classify images into scenes and identify their respective landform types; Step 3: Based on the scene, call the corresponding degradation model parameters to simulate a low-resolution image; Based on different terrains, specific downsampling methods are selected, and after obtaining the noise level and edge enhancement weight, the image is sent to its corresponding degradation model to simulate a low-resolution image. Step 4: Perform initial upsampling using NEDI interpolation; The NEDI method is used to upsample the low-resolution image to complete the interpolation operation, resulting in a high-resolution initial reference frame with the target resolution. Step 5: Iteratively execute the convex set projection reconstruction process; The motion parameter estimation process is used to process each pixel in the low-resolution image one by one. Block matching is used to find the center coordinates of the corresponding PSF window of the corresponding pixel in the high-resolution initial frame. Point spread function correction is performed based on the residual of pixel values ​​between the simulated and real low-resolution images. Step 6: Extract edge priors in each iteration and perform gradient-guided enhancement; Edge detail reconstruction is performed by using the Canny operator to extract image edge priors and Sobel to extract image edge gradients, followed by edge-guided enhancement to preserve structural details and improve perceptual quality. Step 7: Iterate through the process, output the final super-resolution image, and save it; Iterative corrections are performed based on convex set projection theory, and image details are reconstructed through back projection error compensation, resulting in a high-resolution image with enhanced details and clear edges.

Images