Image edge feature enhancement correction fusion method based on guide filter

Abstract

The invention discloses an image edge feature enhancement correction fusion method based on a guide filter, and aims to solve the problems that detail features of a low-illumination visible light image and an infrared image are not obvious, focusing edge information is not clear, and registration of a multi-focus image is wrong. The invention provides an edge feature enhancement correction fusion method based on a guide filter. In the illumination enhancement stage, a visible light image is divided into a base layer and a detail layer through a guide filter, and the image contrast and detail information are enhanced. For an infrared image, an infrared background is reconstructed by using a quadtree decomposition and Bezier interpolation method, unclear focusing edge feature information is extracted, and the visibility of image details is enhanced. And finally, reconstructing the two processed images by using a multi-scale weighted gradient method, and performing fusion by calculating large-scale and small-scale weight scales.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image fusion, and in particular to an image edge feature enhancement correction fusion method based on a guided filter. BACKGROUND

[0002] In the fields of visual monitoring, image tracking, medical imaging and automatic driving, a single type of sensor cannot obtain complete information of a target and a scene, and thus image fusion technology is needed to integrate multi-source images from different sensors into an image with more comprehensive information. An infrared sensor can overcome the influence of light conditions and easily obtain infrared radiation characteristics of a scene, but its size is often too large, its resolution is low and it lacks high-frequency details. A visible light sensor describes background and detail information through light reflection, has higher resolution and detail texture information, and is more in line with human visual perception characteristics, but image quality decreases when light is insufficient in bad weather. Therefore, complementary fusion of infrared source images and visible light source images can obtain more comprehensive scene and better visual effect information. SUMMARY

[0003] The image edge feature enhancement correction fusion method based on a guided filter provided by the present application can at least solve one of the technical problems in the background art.

[0004] To achieve the above-mentioned purpose, the present application adopts the following technical solutions: An image edge feature enhancement correction fusion method based on a guided filter, executed by a computer device, includes the following steps: S100, selecting source images and extracting visible light images and infrared images; S200, pre-processing the visible light images and infrared images of the original images to obtain enhanced images; S300, fusing the reconstructed images of the visible light images and infrared images by calculating scale weights; The method for obtaining enhanced images includes: S210, performing image enhancement processing on the visible light images to normalize the input images I to the range of [0, 255], and performing logarithmic transformation on the normalized images; S211, extracting edge features of the visible light images to construct an edge weight map; S212, performing differential enhancement based on the edge weight map to obtain an optimized base layer and a detail layer; S213, obtaining enhanced visible light images .

[0005] Further, the edge weight map construction method in step S211 of the present application comprises: Structured edge detection: For each pixel p of the visible light image Ivis, a patch P(p) around it is taken, which is represented as a high-dimensional feature vector of a series of features Where θ1 is a feature extraction parameter; The probability of the pixel belonging to an edge is predicted using a pre-trained random forest model; the prediction process is as follows: The feature vector is input into the pre-trained model to obtain a probability value S(p), indicating the probability of the pixel being predicted to belong to an edge; The model outputs a probability value S(p) ∈ [0, 1], indicating the probability of the pixel being an edge; The model formula is as follows: Where Fmodel is the prediction model and θ2 is the model parameter; Construct an initial edge response map: the output result of the prediction model is subjected to SED detection to obtain an edge response map E, each pixel value in E being the probability value S(p) of the prediction model; In order to obtain a transition-smooth weight map and eliminate very small noise responses, the edge response image E is subjected to Gaussian smoothing and then normalized, wherein the Gaussian smoothing method is as follows: ; In the formula, is a Gaussian convolution kernel with a standard deviation of σ, * is a convolution operation, and σ is a control of the smoothing degree, usually taking a value between 1.0 and 5.0, the larger σ is, the smoother the transition in the edge region is;

[0006] The linear mapping of Esmooth to the range [0, 1] is performed by the above formula to obtain the final edge weight map ; Wherein, tends to 1 for an edge region, to 0 for a flat region, and to 0.5 for a texture region.

[0007] Further, the optimized base layer acquisition method in step S212 of the present application comprises: The base layer of the image is extracted using a fast directional filter to obtain low-frequency information of the image;

[0008] In the formula: represents the filtered image, is a filter radius, is the regularization coefficient, and and denote the size of the filter and the edge preserving degree, respectively, =0.01;

[0009] is the extracted base layer image, is the natural logarithm operator, =1 to prevent the logarithm value from being negative; The adaptive filter is set as:

[0010] where epsmin is a preset minimum regularization coefficient, typically 0.001 to 0.01, and epsmax is a preset maximum regularization coefficient, typically 0.1 to 0.5; The adaptive filter is substituted into the fast guided filter calculation to obtain a more clean base layer .

[0011] Further, the method for obtaining the optimized detail layer in step S211 of the present application comprises: The detail layer preserves the high-frequency information of the image, which changes rapidly in detail texture, edge profile, and is represented as follows:

[0012] In the formula, is the detail layer image; Based on the edge weight map An adaptive gain coefficient G is constructed map , and the formula is as follows:

[0013] where Gmin is a preset minimum gain coefficient, typically set to 0.3 to 1.0, and the smaller the value, the stronger the suppression effect on non-edge regions. G max is a preset maximum gain coefficient, typically set to 1.2 to 3.0, and the larger the value, the more exaggerated the enhancement effect on the edge; Then, the detail layer image is compared and optimized to obtain an optimized detail layer image , which is shown as follows:

[0014] In the formula, is element-wise multiplication.

[0015] Further, the method for pre-processing the infrared image in step S200 of the present application comprises: The infrared image is decomposed into a suitable size through a quadtree structure to roughly estimate the outline of the infrared target; Then, 16 control points are uniformly sampled for each block to obtain the corresponding x and y coordinates; The grayscale values ​​of the control points are set to local minimum grayscale values ​​to represent the background. The B-ezier surface of each block is reconstructed using the x-coordinates, y-coordinates, and grayscale values ​​of the 16 control points, as shown in the following equation:

[0016] For interpolation points, the pixel position they represent is indicated by an interpolation ratio from 0 to 1. These are the variable interpolation coefficients related to the interpolation points. The interpolation coefficient matrix is... for Matrix;

[0017]

[0018] when Sometimes,

[0019]

[0020] In the formula Representing feature points Infrared features (x-coordinate, y-coordinate, image intensity). Let x represent the x-coordinate, and Q represent the interpolated x-coordinate; Control points are sampled uniformly within each square block, by uniform setting. and The B-ezier surface of each block is reconstructed, and then the B-ezier surfaces of each block are combined in a quadtree structure to reconstruct the infrared background. A smooth infrared background is obtained using a Gaussian filter.

[0021] Furthermore, the method for fusing the reconstructed images from the visible light image and the infrared image in step S300 of the present invention includes: Multi-scale fusion, i.e., large-scale and small scale Using large-scale focusing measures The focus region of each input image is coarsely detected, and then a small-scale focus metric is used. Determine the gradient weights near the boundary of the focused region; By comparing the large-scale saliency values ​​of each pixel in all input images, the key regions in each image are detected as follows:

[0022] Define an unknown region near the boundary of the focused region to separate the clearly defined focused region and the defocused region. In multi-scale weighted gradient fusion, determine the gradient value of the unknown region at a small scale. Small-scale significance value It is susceptible to noise and other factors, resulting in the capture of many pseudo-focus areas, especially at small scales. The detected pseudo-focused region is defined as:

[0023] Utilizing large-scale focus saliency The focus area is roughly detected, preserving the main structure of the image; By utilizing small-scale saliency, the defined focus area is correctly identified, effectively suppressing errors introduced at the boundary of the focus area and obtaining the texture details of the image; Image reconstruction, for each input image At a given point Its significance can be calculated. The value is assigned to the weight of the saliency of the local image structure. The weighted structure tensor is then calculated as follows:

[0024] Obtain weight for:

[0025] The maximum and minimum eigenvalues ​​were calculated simultaneously. That is, the eigenvectors of the matrix and In the same direction; Multiscale weighted structure tensor based on image saliency The largest eigenvalue derived And the corresponding feature vectors, merging the gradients of each component of the multi-source image, assuming It is the fused image, made by making its gradient as close as possible. To reconstruct the image, we can obtain the image after gradient fusion and reconstruction:

[0026] when At that time, we can obtain: .

[0027] For the reconstructed fused image, To fuse the Laplacian values ​​of the images, For the target gradient field, This represents the divergence operator.

[0028] From the above technical solutions, the application fuses the visible light and infrared images after image enhancement, the fused image has clear texture of detail features and obvious edge features, provides high-quality input images for high-level visual task processing, and can improve accuracy and reduce misidentification rate in visual image processing and automatic driving field target detection and identification task processing. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 The visible light image enhancement schematic diagram of the application; Figure 2 The infrared feature extraction schematic diagram of the application; Figure 3 The fusion algorithm flow schematic diagram of the application. DETAILED DESCRIPTION

[0030] To make the purpose, technical solutions and advantages of the embodiments of the application clearer, the technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are part of the embodiments of the application, rather than all the embodiments of the application.

[0031] As shown in the figure, Figure 1 The image edge feature enhancement correction fusion method based on a guided filter includes the following steps: S100, selecting a source image and extracting a visible light image and an infrared image; S200, pre-processing the visible light image and the infrared image of the source image to obtain an enhanced image; S300, fusing the visible light image and the infrared image enhanced image by calculating a scale weight.

[0032] The steps will be described in detail as follows: S100, selecting a source image and extracting a visible light image and an infrared image, for the same scene, collecting infrared images and visible light images by using different cameras, extracting bright features of the visible light image and infrared features of the infrared image, and fusing them into an image to enhance the visibility of the image.

[0033] S200, pre-processing the visible light image and the infrared image of the source image to obtain an enhanced image; S210, performing image enhancement processing on the visible light image to normalize it to an input image I in the range of [0, 255], and performing logarithmic transformation on the normalized image; S211, extracting edge features of the visible light image to construct an edge weight map; The edge weight map construction method is as follows: Structured edge detection: For each pixel p of the visible light image I vis , a patch P(p) around the pixel is taken, and the patch around the pixel is represented as a high-dimensional feature vector of a series of features , where θ1 is a feature extraction parameter.

[0034] A pre-trained model (random forest) is used to predict the probability that the pixel belongs to an edge; the prediction process is as follows: The feature vector is input into the pre-trained model to obtain a probability value S(p), indicating the probability that the pixel being predicted belongs to an edge. The model outputs a probability value S(p) ∈ [0, 1], indicating the probability that the pixel is an edge.

[0035] The model formula is as follows: Where F model is the prediction model, and θ2 is the model parameter.

[0036] Construct an initial edge response map: the output of the prediction model is detected by SED to obtain an edge response map E, and each pixel value in E is the probability value S(p) of the prediction model; In order to obtain a transition smooth weight map and eliminate very small noise responses, the edge response image E is Gaussian smoothed and then normalized, and the Gaussian smoothing method is as follows:

[0037] In the formula, is a Gaussian convolution kernel with a standard deviation of σ, * is a convolution operation, and σ is a control of the smoothing degree, usually taking a value between 1.0 and 5.0, and the larger the σ, the smoother the transition in the edge region.

[0038]

[0039] Through the above formula, E smooth is linearly mapped to the range [0, 1] to obtain the final edge weight map .

[0040] Where, tends to 1 for edge regions, to 0 for flat regions, and to 0.5 for texture regions.

[0041] S212, based on the edge weight map, differential enhancement is performed to obtain an optimized base layer and a detail layer; The base layer of the image is extracted by using the fast guided filter to obtain the low-frequency information of the image, such as the basic information of the overall structure, the general contour, etc.

[0042]

[0043] The fastguidedfilter represents the fast guided filter calculation, which is used for image filtering or up-sampling while maintaining edge smoothness. In the formula: represents the filtered image, is the filter radius, is the regularization coefficient, and respectively represents the size of the filter and the edge preservation degree.

[0044]

[0045] is the extracted base layer image, is the natural logarithm operator, =1 to prevent the logarithmic value from being negative.

[0046] The detail layer retains the high-frequency information of the image, such as the detail texture, edge contour, etc., which changes rapidly, and is represented as follows:

[0047] In the formula: is the detail layer image.

[0048] Set adaptive filtering:

[0049] Where epsmin is the preset minimum regularization coefficient, usually 0.001 to 0.01, and epsmax is the preset maximum regularization coefficient, usually 0.1 to 0.5.

[0050] Small eps is used in the edge area to maintain the edge, and large eps is used in the flat area to force smoothness to complete the guided filtering.

[0051] The adaptive filtering is substituted into the fast guided filter calculation to obtain the optimized and cleaner base layer .

[0052] Based on the edge weight map An adaptive gain coefficient G is constructed map , and the formula is as follows:

[0053] Where Gmin is the preset minimum gain coefficient, usually set between 0.3 and 1.0, and the smaller the value, the stronger the suppression effect on the non-edge area. maxThe preset maximum gain coefficient is usually set between 1.2 and 3.0, and the greater the value, the more exaggerated the enhancement effect on the edge.

[0054] Then, the detail layer image is compared and optimized to obtain an optimized detail layer image , as follows:

[0055] In the formula, is element-wise multiplication.

[0056] S213, obtaining an enhanced visible light image according to the low-frequency information and the high-frequency information of the visible light image .

[0057] Through the scale factor dynamic range compression is performed on the bottom layer, to restore the overall contrast.

[0058]

[0059] Finally, the enhanced image is obtained as:

[0060] S220, processing the infrared image to extract infrared features to obtain detail and edge information of the target image; First, the infrared image is decomposed into a suitable size through a quadtree structure to roughly estimate the outline of the infrared target. Then, 16 control points are uniformly sampled for each block to obtain the corresponding x coordinate and y coordinate. Since the quadtree structure can confirm most of the vertices on the infrared background, but there may still be a few points inside the object in the infrared image, the gray value of the control point is set to the local minimum gray value to represent the background. The B-ezier surface of each block is reconstructed through the x coordinate, y coordinate and gray value of the 16 control points, as follows:

[0061] is an interpolation point, and the pixel position represented by the interpolation point is represented by an interpolation ratio of 0 to 1, is a variable interpolation coefficient related to the interpolation point, is an interpolation coefficient matrix, is the matrix.

[0062]

[0063]

[0064] When , there is

[0065]

[0066] wherein denotes the feature points (x coordinate, y coordinate, image intensity), denotes the x coordinate, and Q denotes the interpolated x coordinate.

[0067] Since the control points are uniformly sampled in each square block, by uniformly setting and , the B-ezier surface of each block is reconstructed, and then the B-ezier surfaces of each block are combined in a quadtree structure to reconstruct the infrared background. However, using different control points in the stitched surface, all the B-ezier surfaces may directly combine to produce blocking artifacts, so a Gaussian filter is used to obtain a smooth infrared background.

[0068]

[0069] denotes the extracted infrared feature image, denotes the original image, is a Gaussian function, representing a filter of s x s, is a parameter for controlling the smoothness.

[0070] Image fusion based on gradient fusion of multi-value images is based on capturing the geometric shape of an image through the gradients of different images, and the gradient of a multi-value image is expressed as:

[0071] The norm expression is:

[0072] The covariance matrix G is also a structure tensor, and its expression is:

[0073] S300, fusing the reconstructed images of the visible light image and the infrared image by calculating the scale weight; A fusion method based on multi-scale weighted gradient is proposed, which introduces the perceptual contrast to reflect the structural significance of the local image, and effectively identifies the local structure of the image by fusing the gradients of different images, and renders it to the fusion image.

[0074] Multi-scale fusion, i.e., large-scale and small-scale , uses a large-scale focus measure The focus region of each input image is coarsely detected, and then a small-scale focus metric is used. Determine the gradient weights near the boundary of the focused region.

[0075] By comparing the large-scale saliency values ​​of each pixel in all input images, the key regions in each image are roughly detected as follows:

[0076] Large-scale focusing measurements can effectively reduce the effects of anisotropic blurring and registration errors, but they are prone to introducing errors near the boundary of the focusing region. Therefore, an unknown region is defined near the boundary of the focusing region to separate the clearly defined focusing region and the defocused region. In multi-scale weighted gradient fusion, the gradient value of the unknown region is determined at a small scale.

[0077] Small-scale significance value It is susceptible to noise and other factors, resulting in the capture of many pseudo-focus areas, especially at small scales. The detected pseudo-focused region is defined as:

[0078] Utilizing large-scale focus saliency The system performs a coarse detection of the focal region, preserving the main structure of the image. Then, by utilizing small-scale saliency, it correctly identifies the defined focal region, effectively suppressing the error introduced at the boundary of the focal region, and obtaining the texture details of the image.

[0079] Image reconstruction, for each input image At a given point Its significance can be calculated. The value is assigned to the weight of the saliency of the local image structure. The weighted structure tensor is then calculated as follows:

[0080] Obtain weight for:

[0081] Simultaneously, the maximum and minimum eigenvalues ​​can be calculated as follows: Therefore, the eigenvectors of the matrix and In the same direction, the multi-scale weighted structure tensor based on image saliency The largest eigenvalue derived And the corresponding feature vectors, merging the gradients of each component of the multi-source image, assuming It is the fused image, made by making its gradient as close as possible. To reconstruct the image, we can obtain the image after gradient fusion and reconstruction:

[0082] When , we can get:

[0083] For the reconstructed fusion image, is the Laplacian value of the fusion image, is the target gradient field, represents the divergence operator.

[0084] The purpose of the present application is to provide an infrared and visible image fusion method based on guided filtering, one is the pre-processing process of the original image information, and the other is the image fusion reconstruction process. The image preprocessing is to enhance the visible light brightness, increase the overall contrast of the image, highlight the detail feature texture information and edge feature details of the infrared image. The image fusion process is a multi-scale weighted gradient method, which reconstructs the two processed images, and fuses them by calculating the weight scale of large and small scales.

[0085] In still another aspect, the present application also discloses a computer device, comprising a memory and a processor, the memory stores a computer program, and the computer program is executed by the processor to make the processor execute the steps of the above method.

[0086] In yet another embodiment provided in the present application, a computer program product containing instructions is also provided, which, when running on a computer, causes the computer to execute any of the above image edge feature enhancement correction fusion methods based on a guided filter.

[0087] It can be understood that the system, device and storage medium provided by the embodiments of the present application correspond to the method provided by the embodiments of the present application, and the explanation, examples and beneficial effects of the related content can refer to the corresponding part in the above method.

[0088] In the embodiments described above, all or some of the steps can be implemented by hardware, software, firmware or any combination thereof. When implemented by software, all or some of the steps can be implemented in the form of one or more computer programs or program elements. The computer programs reside (at least temporarily) in a memory of a computer during execution. The memory can be a RAM memory, a flash memory, a ROM memory, an EPROM memory, or any other suitable memory. The memory can be integral to or separate from the computer. The computer programs can be written in any suitable programming language, such as C, C++, Java, Visual Basic, etc. The computer programs can be written in assembly or machine language, if desired. The computer programs can be distributed over network coupled file servers, or can be distributed by any other suitable means.

[0089] It is to be noted that, in the present document, relational terms such as first and second and the like can be used solely to distinguish one entity or action from another entity or action without necessarily implying any actual relationship or order between such entities or actions. Also, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. In addition, terms such as "first" and "second" are used herein only to distinguish one element from another, without necessarily implying any actual relationship or order between such elements.

[0090] Each of the embodiments described in the present specification is described in a related manner, and the same or similar parts between the embodiments can be referred to each other. Each of the embodiments mainly describes the difference from other embodiments. In particular, the system embodiments are described simply because they are basically similar to the method embodiments, and the same or similar parts can be referred to the description of the method embodiments.

[0091] The above examples are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some technical features can be replaced by equivalent features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for enhancing, correcting, and fusing image edge features based on guided filters, characterized in that, Perform the following steps using a computer device: S100: Select the source image and extract the visible light and infrared images; S200: Preprocess the visible light and infrared images of the source image to obtain the enhanced image; S300: The enhanced images of visible light and infrared images are fused by calculating scale weights; The methods for obtaining the enhanced image include: S210, Visible light image Image enhancement processing is performed, and the image is normalized to the input image I within the range of [0,255]. Logarithmic transformation is then applied to the normalized image. S211. Extract edge features from visible light images and construct an edge weight map; S212. Perform differential enhancement based on the edge weight map to obtain the optimized base layer and detail layer; S213. Obtain the enhanced visible light image based on the optimized base layer and detail layer. .

2. The image edge feature enhancement, correction, and fusion method based on a guided filter according to claim 1, characterized in that, The edge weight graph construction method in step S211 includes: Structured edge detection: For each pixel p in the visible light image Ivis, take a patch P(p) around the pixel, and represent the patch around the pixel as a high-dimensional feature vector of a series of features. , where θ1 is the feature extraction parameter; The probability of a pixel belonging to an edge is predicted using a pre-trained random forest model; the prediction process is as follows: eigenvectors Input the pre-trained model and obtain the probability value S(p). This indicates the probability that the pixel being predicted belongs to an edge; The model outputs a probability value S(p) ∈ [0, 1], which represents the probability that the pixel being predicted is an edge; The model formula is as follows: Where Fmodel is the prediction model, and θ2 are the model parameters; Constructing the initial edge response map: The output of the prediction model is detected by SED to obtain the edge response map E, where each pixel value in E is the probability value S(p) of the prediction model. To obtain a smooth weight map and eliminate minimal noise responses, the edge response image E is Gaussian smoothed and then normalized. The Gaussian smoothing method is as follows: ； In the formula, It is a Gaussian convolution kernel with a standard deviation of σ, * represents the convolution operation, and σ controls the smoothness. Normalize using the above formula, and then... smooth A linear mapping is applied to the range [0, 1] to obtain the final edge weight map. ; in, A tendency of 1 indicates an edge region, a tendency of 0 indicates a flat region, and a tendency of 0.5 indicates a textured region.

3. The image edge feature enhancement, correction, and fusion method based on guided filters according to claim 1, characterized in that, The optimized method for obtaining the base layer in step S212 includes: The base layer of the image is extracted using a fast guided filter to obtain the low-frequency information of the image; In the formula: This represents the filtered image. The filter radius is... The regularization coefficients represent the filter size and the degree of edge preservation, respectively. =0.01; For the extracted base layer image, For the natural logarithm operator, =1 prevents the logarithm from being negative; Configure adaptive filtering: Where epsmin is the preset minimum regularization coefficient and epsmax is the preset maximum regularization coefficient; Substituting adaptive filtering into the fast guided filtering calculation, we obtain a cleaner, optimized base layer. .

4. The image edge feature enhancement, correction, and fusion method based on a guided filter according to claim 3, characterized in that, The optimized detail layer acquisition method in step S211 includes: The detail layer preserves high-frequency information about the image's detailed textures and rapidly changing edge contours, and is represented as follows: In the formula: For detail layer images; Based on edge weight graph Constructing adaptive gain coefficients G map The formula is as follows: Where Gmin is the preset minimum gain coefficient, G max The preset maximum gain coefficient; Then, the image is compared and optimized with the detail layer image to obtain the optimized detail layer image. The announcement is as follows: In the formula, This is for element-wise multiplication.

5. The image edge feature enhancement, correction, and fusion method based on a guided filter according to claim 1, characterized in that, The infrared image preprocessing method in step S200 includes: The infrared image is decomposed into appropriate sizes using a quadtree structure to roughly estimate the outline of the infrared target. Then, 16 control points are uniformly sampled for each block to obtain the corresponding x and y coordinates; The grayscale values ​​of the control points are set to local minimum grayscale values ​​to represent the background. The B-ezier surface of each block is reconstructed using the x-coordinates, y-coordinates, and grayscale values ​​of the 16 control points, as shown in the following equation: For interpolation points, the pixel position they represent is indicated by an interpolation ratio from 0 to 1. These are the variable interpolation coefficients related to the interpolation points. The interpolation coefficient matrix is... for Matrix; when Sometimes, In the formula Representing feature points infrared characteristics, Let x represent the x-coordinate, and Q represent the interpolated x-coordinate; Control points are sampled uniformly within each square block, by uniform setting. and The B-ezier surface of each block is reconstructed, and then the B-ezier surfaces of each block are combined in a quadtree structure to reconstruct the infrared background. A smooth infrared background is obtained using a Gaussian filter.

6. The image edge feature enhancement, correction, and fusion method based on a guided filter according to claim 1, characterized in that, The method for fusing the reconstructed images from the visible light image and the infrared image in step S300 includes: Multi-scale fusion, i.e., large-scale and small scale Using large-scale focusing measures The focus region of each input image is coarsely detected, and then a small-scale focus metric is used. Determine the gradient weights near the boundary of the focused region; By comparing the large-scale saliency values ​​of each pixel in all input images, the key regions in each image are detected as follows: Define an unknown region near the boundary of the focused region to separate the clearly defined focused region and the defocused region. In multi-scale weighted gradient fusion, determine the gradient value of the unknown region at a small scale. Small-scale significance value It is susceptible to noise and other factors, resulting in the capture of many pseudo-focus areas, especially at small scales. The detected pseudo-focused region is defined as: Utilizing large-scale focus saliency The focus area is roughly detected, preserving the main structure of the image; By utilizing small-scale saliency, the defined focus area is correctly identified, effectively suppressing errors introduced at the boundary of the focus area and obtaining the texture details of the image; Image reconstruction, for each input image At a given point Its significance can be calculated. The value is assigned to the weight of the saliency of the local image structure. The weighted structure tensor is then calculated as follows: Obtain weight for: The maximum and minimum eigenvalues ​​were calculated simultaneously. That is, the eigenvectors of the matrix and In the same direction; Multiscale weighted structure tensor based on image saliency The largest eigenvalue derived And the corresponding feature vectors, merging the gradients of each component of the multi-source image, assuming It is the fused image, made by making its gradient as close as possible. To reconstruct the image, we can obtain the image after gradient fusion and reconstruction: when At that time, we can obtain: 。 For the reconstructed fused image, To fuse the Laplacian values ​​of the images, For the target gradient field, This represents the divergence operator.

Images