Image enhancement method and system based on deep learning

Abstract

The invention discloses an image enhancement method and system based on deep learning, and the method comprises the following steps: obtaining an image pixel matrix, extracting multi-region features according to a brightness difference value, a gradient direction and a texture density partition frequency, carrying out the weighting of the multi-region features, generating a response graph, carrying out the scale decomposition, extracting multi-scale features, fusing a residual image, and enhancing the detail inhibition redundancy. And calling the multi-image parameter to callback the brightness, and outputting an enhanced image. According to the method, accurate separation of high and low frequency regions is realized through region frequency attribute division, fuzzy and excessive enhancement caused by unified processing are avoided, a differentiation strategy is adopted for different regions to improve high-frequency details and suppress low-frequency redundancy, and feature response weight maps are subjected to weighted stacking and normalized enhancement of local textures and contrast. Scale decomposition is combined with multi-scale fusion to improve naturalness and avoid visual discomfort, and edge consistency matching optimizes brightness callback to enhance edge definition and overall visual performance.

Description

Technical Field

[0001] The present invention relates to the field of non-spatial domain filtering technology, and in particular to an image enhancement method and system based on deep learning. Background Art

[0002] The field of non-spatial domain filtering technology includes technical methods for improving image quality during image processing, especially image enhancement methods. This technical field mainly involves improving the contrast, brightness, clarity and other visual effects of images by processing the non-spatial domain features of images. Non-spatial domain filtering technology usually does not rely on the spatial pixel values ​​of the image, but optimizes by analyzing the frequency domain information of the image or the characteristics of other transform domains. In this field, the frequency components, color models, time domain features or other mathematical transformations of the image are usually used as the basis for filtering. This technical field is not only widely used in image enhancement, but also has important significance in the fields of medical imaging, satellite image processing, industrial automation, etc.

[0003] Among them, the image enhancement method refers to improving the image through a specific processing method, thereby improving the visual effect of the image. This method targets the problem of decreased visual quality in the image due to noise, poor lighting conditions, insufficient contrast, etc. The image enhancement method is usually implemented by adjusting the grayscale value of the image, optimizing the contrast, enhancing the color, etc. Specifically, the patent subject performs frequency domain filtering on the image in the non-spatial domain and uses frequency component analysis to distinguish image features in different areas. The purpose of enhancing image details is achieved by processing high-frequency and low-frequency information differently. The patented method uses frequency domain transformation technology, such as Fourier transform or wavelet transform, to convert the image from the spatial domain to the frequency domain, and then optimizes the image performance by selectively filtering and enhancing specific frequency components.

[0004] The existing technology lacks the fine division and dynamic adjustment of image area characteristics, and often uniformly processes the frequency components of the entire image, resulting in inaccurate processing of image details and redundant information. In low-frequency areas, the noise suppression effect is not ideal, while high-frequency areas are prone to over-enhancement, resulting in distorted or unnatural details. Traditional methods are relatively simple in processing edge areas, which are prone to blurring and affect the overall image clarity. In practical applications, especially in the fields of medical and satellite images, insufficient or distorted image details can lead to the loss of key details and affect the accuracy of the analysis results. These deficiencies come from the fact that the existing technology fails to fully consider regional differences and lacks more flexible means of detail enhancement. Summary of the invention

[0005] The purpose of the present invention is to solve the shortcomings of the prior art and propose an image enhancement method and system based on deep learning.

[0006] In order to achieve the above object, the present invention adopts the following technical solution: an image enhancement method based on deep learning, comprising the following steps: S1: Get the pixel matrix of the input image, determine the regional frequency attribute according to the brightness difference, gradient direction change value and local texture density, divide it into high-frequency area and low-frequency area according to the edge response threshold, and obtain the image frequency partition label; S2: extracting edge intensity values, texture fineness values ​​and contrast values ​​of multiple regions based on the image frequency partition labels, performing normalized weighted superposition operations, assigning the results to corresponding regions, and obtaining a regional feature response weight map; S3: According to the regional feature response weight map, the input image is scaled and decomposed, a multi-scale edge-guided feature map, a texture map and a brightness map are extracted, the multi-map residual response value is adjusted according to the weight, the high-frequency regional details are enhanced and the low-frequency regional redundancy is suppressed, all scale maps are fused and a mapping structure is constructed to obtain an image enhancement residual map; S4: Based on the image enhancement residual map, call the scale map brightness mean, edge direction distribution map and texture contrast map, match the area according to the edge consistency coefficient and perform brightness callback to obtain a multi-scale enhanced fusion image.

[0007] As a further solution of the present invention, the image frequency partition label is specifically a high-frequency area and a low-frequency area, the regional feature response weight map includes multi-region edge intensity values, texture fineness values, and contrast values, the image enhancement residual mapping map is specifically a multi-scale edge-guided feature map, texture mapping map, and brightness mapping map after scale decomposition, and the multi-scale enhanced fusion image includes a scale map brightness mean, an edge direction distribution map, and a texture contrast map.

[0008] As a further solution of the present invention, the step of obtaining the image frequency partition label is specifically as follows: S101: Based on the input image pixel matrix, the grayscale difference is calculated and compared with the brightness difference threshold, the number of pixels with significant brightness difference is counted and normalized, and the brightness difference density value is obtained; S102: calling the brightness difference density value, combining the multi-pixel gradient direction change difference, the neighborhood gradient angle difference, and the local change trend intensity, using the formula: ; The regional gradient response value is obtained by calculation and superimposed with the distribution of brightness difference density to obtain the frequency mutation factor value; in, Represents the frequency mutation factor value, Represents the gradient direction change difference of the a-th pixel pair, Indicates the trend strength of the local gradient direction change at the point, is the point gradient change density value, is the brightness difference density value of the point neighborhood, Indicates the number of pixel pairs participating in the statistics; S103: comparing the frequency mutation factor value with the texture density distribution and the edge response threshold, determining the regional frequency attribute, and generating an image frequency partition label.

[0009] As a further solution of the present invention, the step of obtaining the regional feature response weight map is specifically: S201: extracting edge intensity values, texture fineness values, and contrast values ​​of multiple regions based on the image frequency partition labels, normalizing them according to the maximum and minimum values, and generating multi-region standardized image feature values; S202: calling the multi-region standardized image feature values, collecting multi-region edge intensity values, texture fineness values, contrast values, gradient direction change values, spectrum balance values ​​and color distribution concentration, performing nonlinear cross integration operations, using the formula: ; Calculate and obtain regional characteristic response values, adjust the attribute response ratio within the region, and generate regional characteristic response distribution values; in, represents the characteristic response value of the bth region, represents the normalized edge strength value of the bth region, represents the normalized texture fineness value of the bth region, represents the normalized contrast value of the bth region, Indicates the gradient direction change value of the b-th region, represents the spectrum balance value of the bth area, represents the color distribution concentration of the bth area, Represents the joint summation of spectrum and color; S203: According to the regional feature response distribution values, a multi-region response value grid is mapped to an image space position, a regional weight distribution matrix is ​​established, and a regional feature response weight map is obtained.

[0010] As a further solution of the present invention, the step of obtaining the image enhancement residual map is specifically as follows: S301: Decomposing the input image on a scale according to the regional feature response weight map, extracting the edge-guided feature map, the texture map and the brightness map, performing regional feature response weight comparison, extracting high-frequency regional features, and generating multi-scale regional response feature values; S302: Call the multi-scale regional response characteristic value, calculate the response difference value, construct the characteristic sensitivity coefficient, and adjust the residual response with the brightness map difference value, using the formula: ; Calculate and obtain the weight-adjusted residual structure; in, represents the weight-adjusted residual structure, represents the intensity difference of the edge-guided feature map at the cth scale, represents the structural difference of the texture map at the cth scale, Represents the grayscale distribution value of the brightness map at the cth scale, represents the characteristic sensitivity coefficient, Indicates the maximum value operation among all scales; S303: adjusting the residual structure amount according to the weight, fusing the residual signals of all scale maps, constructing a multi-scale structural relationship of image elements, and generating an image enhancement residual map.

[0011] As a further solution of the present invention, the step of acquiring the multi-scale enhanced fusion image is specifically as follows: S401: Based on the image enhancement residual map, call the scale map brightness mean and edge direction distribution map, calculate the brightness offset difference and the edge direction angle value, select the area that meets the edge consistency threshold condition, and obtain the edge consistency area distribution coefficient; S402: Extract the texture contrast value of the matching area according to the edge consistent area distribution coefficient, and construct the brightness difference and texture contrast difference ratio using the formula: ; Calculate the ratio of overall brightness deviation to texture distinctness, reorder the regional brightness according to texture features, and obtain a brightness callback reference sequence; in, Represents the ratio of overall brightness deviation to texture sharpness. Indicates the number of pixels involved in the calculation in the edge consistent area. Represents the brightness value of the dth pixel in the edge consistent area, represents the mean brightness of the pixel in the scale map, Indicates the texture contrast value corresponding to the d-th pixel in the texture contrast map, Indicates the maximum value of the texture contrast value corresponding to all participating pixels in the current area. Represents the brightness difference between the dth pixel and the adjacent rth pixel; S403: Based on the brightness callback reference sequence, call the brightness information of the matching area, perform brightness mapping and segmented adjustment, integrate texture features and edge direction features, and obtain a multi-scale enhanced fused image.

[0012] As a further embodiment of the present invention, the method further comprises: S5: According to the multi-scale enhanced fusion image, the edge response difference map, texture consistency offset value and brightness change rate of the original image and the enhanced image are called, and the enhancement amplitude judgment operation is performed on all high-frequency areas in the image, and the area with a response amplitude lower than the edge enhancement judgment threshold is eliminated, and the original image value is restored to it, and the regional enhanced image result map is output; The region enhanced image result map is specifically an image after the high-frequency region enhancement amplitude is determined.

[0013] As a further solution of the present invention, the step of obtaining the regional enhanced image result map is specifically: S501: According to the multi-scale enhanced fusion image, call the edge response difference map of the original image and the enhanced image to obtain the regional edge response strength, compare the gradient response value of the difference map at the regional edge position with the set edge enhancement discrimination threshold, and determine whether the region belongs to the high-frequency change region, and generate a high-frequency response interval value; S502: calling the high frequency response interval value, combining the texture consistency offset value and the brightness change rate, and using the formula for all areas in the high frequency response interval: ; Calculate the enhancement amplitude index, filter the area greater than the edge enhancement discrimination threshold, and obtain the edge enhancement effective coefficient distribution; in, Represents the enhancement amplitude index of the coordinate (e, s) position, The texture consistency offset value representing the position, represents the rate of change of brightness, represents the response value of the edge response difference map, represents the enhancement value of the x-th layer fusion image at the position under the differentiated scale, and v represents the number of scales involved in the fusion image; S503: calling the edge enhancement effective coefficient distribution, identifying the area below the edge enhancement discrimination threshold, restoring the image value to the original image value, and maintaining the enhancement process for the remaining areas to obtain the regional enhanced image result map.

[0014] An image enhancement system based on deep learning, wherein the image enhancement system based on deep learning is used to execute the above-mentioned image enhancement method based on deep learning, and the system comprises: The image frequency division module obtains the input image pixel matrix, extracts the brightness difference, gradient direction change value and local texture density value, makes judgments based on the brightness difference and density product and the gradient direction change value, divides the frequency area and marks the image block type, and generates the image frequency partition label; The regional feature mapping module calls the image frequency partition label, extracts the regional edge intensity value, texture fineness value and contrast value, normalizes them and weightedly superimposes and assigns them to the corresponding image block position to generate a regional feature response weight map; The scale residual fusion module calls the regional feature response weight map, obtains the image block content and extracts the multi-scale edge-guided map, texture map and brightness map, performs residual calculation and weighted fusion on the three maps according to the image block weight value, and reconstructs the whole image by superimposing the fusion value according to the image block position to generate an image enhancement residual map; The edge brightness callback module calls the image enhancement residual map, collects the brightness mean, edge direction distribution map and texture contrast map, determines the regional matching relationship according to the edge direction consistency coefficient and performs brightness linear callback to generate a multi-scale enhanced fusion image; The high-frequency detail restoration module calls the multi-scale enhanced fusion image, extracts the edge response difference map between it and the original image, and makes a combined judgment based on the texture consistency offset value and the brightness change rate, judges that the unsatisfactory area is restored to the original image value, and generates a regional enhanced image result map.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are: In the present invention, accurate separation of high-frequency and low-frequency regions is achieved by dividing regional frequency attributes, avoiding the blurring and over-enhancement problems caused by overall unified processing. Different processing strategies are adopted for different regions to effectively enhance the details of high-frequency regions and suppress redundant information in low-frequency regions. The feature response weight map enhances the local texture and contrast in the image through weighted superposition and normalization processing, making the details and structure of the image more prominent. In addition, the scale decomposition method is combined with the fusion of multi-scale maps to make the image enhancement effect more natural and avoid visual discomfort caused by over-processing. The brightness callback is optimized through edge consistency matching, making the details of the edge area clearer, the overall visual effect of the image richer, and enhancing the expressiveness and readability of the image. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic diagram of the workflow of the present invention; Figure 2 A flow chart of the steps for obtaining the image frequency partition label of the present invention; Figure 3 A flow chart of the steps for obtaining a regional characteristic response weight map of the present invention; Figure 4 A flowchart of the steps for obtaining an image enhancement residual map according to the present invention; Figure 5 This is a flow chart of the steps for acquiring a multi-scale enhanced fusion image of the present invention; Figure 6The present invention is a flowchart of the steps for obtaining the regional enhanced image result map. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0018] In the description of the present invention, it should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, in the description of the present invention, "multiple" means two or more, unless otherwise clearly and specifically defined.

[0019] Embodiment 1: See also Figure 1 The present invention provides a technical solution: an image enhancement method based on deep learning, comprising the following steps: S1: Get the pixel matrix of the input image, determine the regional frequency attribute according to the brightness difference, gradient direction change value and local texture density, divide it into high-frequency area and low-frequency area according to the edge response threshold, and obtain the image frequency partition label; S2: Based on the image frequency partition label, extract the edge intensity value, texture fineness value and contrast value of multiple regions, perform normalized weighted superposition operation, assign the result to the corresponding region, and obtain the regional feature response weight map; S3: According to the regional feature response weight map, the input image is scaled and decomposed, and the multi-scale edge-guided feature map, texture map and brightness map are extracted. The residual response value of the multi-map is adjusted according to the weight to enhance the details of the high-frequency area and suppress the redundancy of the low-frequency area. All scale maps are fused and a mapping structure is constructed to obtain an image enhancement residual map. S4: Based on the image enhancement residual map, call the scale map brightness mean, edge direction distribution map and texture contrast map, match the area according to the edge consistency coefficient and perform brightness callback to obtain a multi-scale enhanced fusion image; S5: Based on the multi-scale enhanced fusion image, the edge response difference map, texture consistency offset value and brightness change rate of the original image and the enhanced image are called, and the enhancement amplitude judgment operation is performed on all high-frequency areas in the image. The area with a response amplitude lower than the edge enhancement judgment threshold is eliminated, and the original image value is restored to it, and the regional enhanced image result map is output.

[0020] The image frequency partition labels are specifically high-frequency areas and low-frequency areas. The regional feature response weight map includes multi-region edge strength values, texture fineness values, and contrast values. The image enhancement residual mapping map is specifically a multi-scale edge-guided feature map, texture mapping map, and brightness mapping map after scale decomposition. The multi-scale enhanced fusion image includes the scale map brightness mean, edge direction distribution map, and texture contrast map. The regional enhanced image result map is specifically an image after the high-frequency region enhancement amplitude is judged.

[0021] See also Figure 2 , the steps to obtain the image frequency partition label are as follows: S101: Based on the input image pixel matrix, the grayscale difference is calculated and compared with the brightness difference threshold, the number of pixels with significant brightness difference is counted and normalized, and the brightness difference density value is obtained; First, read the image pixels row by row and column by column, and calculate the gray value of each pixel using the standard RGB weighted average formula (Gray=0.299R+0.587G+0.114B) to obtain the gray value of each pixel, and then establish a gray image matrix. For example, for a 4×4 resolution image pixel matrix (Table 1 lists the RGB value of each pixel and the corresponding gray calculation results):

[0022] As shown in Table 1, the grayscale difference operation is performed one by one for adjacent pixels in the image grayscale matrix. For example, the grayscale difference of pixels at positions (1,1) and (1,2) is calculated as: , Similarly, the grayscale difference between (1,3) and (1,4) pixels is calculated as: , execute the above process for all adjacent pixel pairs one by one and obtain all grayscale difference values. Next, compare the grayscale difference results of each pair of pixels with the brightness difference threshold one by one, where the brightness difference threshold is set in the interval [0,255] according to empirical data. Here, the threshold is set to 10. The reason is that in actual scenes, when the brightness difference threshold is set to 10, it can more accurately distinguish between areas with obvious brightness changes and areas with insignificant changes. For example: the grayscale difference of the above (1,1)-(1,2) pixels is 4, which is less than the threshold of 10 and is recorded as a non-significant difference point, while the grayscale difference of (1,3)-(1,4) is 3, which is also recorded as a non-significant difference point. If the grayscale difference of a certain point is 15, it is recorded as a significant difference point, and all adjacent point pairs are judged one by one. Then, the number of all pixels with significant brightness differences in the entire image is counted and normalized by dividing it by the total number of pixel pairs in the image. For example, for an image containing 500 pixel pairs, the number of points with significant brightness differences is 120, and the brightness difference density value is 120 / 500=0.24, which is the "brightness difference density value" obtained in this step.

[0023] S102: Call the brightness difference density value, combine the multi-pixel gradient direction change difference, the neighborhood gradient angle difference, and the local change trend intensity, and use the formula: ; The regional gradient response value is obtained by calculation and superimposed with the distribution of brightness difference density to obtain the frequency mutation factor value; in, Represents the frequency mutation factor value, Represents the gradient direction change difference of the a-th pixel pair, Indicates the trend strength of the local gradient direction change at the point, is the point gradient change density value, is the brightness difference density value of the point neighborhood, Indicates the number of pixel pairs participating in the statistics; Based on the "brightness difference density value" obtained in the previous step, the gradient direction change difference of multiple pixel points in the image is collected. , the intensity of the change trend of the local gradient direction , Gradient change density value And the neighborhood brightness difference density value The specific process of obtaining the above parameters is as follows: For example, three groups of pixel pairs are randomly selected in the image pixel area as an example, and the data is as follows: (Pixel pair 1: , , , ), (pixel pair 2: , , , ), (pixel pair 3: , , , ). Substitute the above data into the formula to calculate the frequency mutation factor value : ; The specific process of step-by-step calculation is as follows: Calculation of pixel pair 1: ; Calculation of pixel pair 2: ; Calculation of pixel pair 3: ; Substituting the sum into the equation, we get: ; Subsequently, the frequency mutation factor value is superimposed on the brightness difference density value, and the final frequency mutation factor value (7.643×0.24=1.834 in this example) is obtained through numerical calculation (such as multiplication). The result is the "frequency mutation factor value" obtained in this step.

[0024] S103: comparing the edge response threshold with the frequency mutation factor value and the texture density distribution, determining the frequency attribute of the region, and generating an image frequency partition label.

[0025] For example, sampling is performed on a local area, different areas are selected, and their texture density values ​​are calculated. The data obtained is assumed to be: [1.0, 1.5, 2.0, 1.8, 1.2], and the average value is: (1.0 + 1.5 + 2.0 + 1.8 + 1.2) / 5 = 1.5, which is used as the texture density reference value. Then, by comparing the frequency mutation factor value 1.834 with the texture density reference value 1.5, the difference between the two is 1.834-1.5 = 0.334, and then compared with the edge response threshold, where the edge response threshold is set to 0.3 (in actual applications, it is generally set between [0.1, 0.5]). In this example, the difference 0.334 is greater than the set threshold 0.3, so the area is judged to be a high-frequency attribute area, and then based on this judgment result, the area is marked as a high-frequency area (labeled as "high frequency"), if it is less than the threshold, it is marked as "low frequency", and the intermediate case is marked as "medium frequency", and finally the frequency attribute division of the entire image area is realized one by one to obtain the frequency partition label.

[0026] See also Figure 3 , the steps for obtaining the regional feature response weight map are as follows: S201: extracting edge intensity values, texture fineness values, and contrast values ​​of multiple regions based on image frequency partition labels, normalizing them according to maximum and minimum values, and generating multi-region standardized image feature values; For example, taking a remote sensing image as the actual object, assume that the image is divided into several feature areas, where area 1, area 2, and area 3 are selected as example study areas. In each area, the initial extraction of edge strength value, texture fineness value, and contrast value is performed. Specifically, for the edge strength value, the Sobel operator is first used to extract edge information in the horizontal and vertical directions respectively. Assuming that the horizontal gradient value and the vertical gradient value obtained in area 1 are 6 and 8 respectively, the edge strength value is calculated as the square root of the sum of the two squares: Then, the texture fineness value in the region is extracted through the gray level co-occurrence matrix. Assuming that the texture fineness of region 1 is initially calculated to be 0.65, the contrast is obtained by the difference between the maximum gray value and the minimum gray value of the pixels in the region. The maximum gray value of the pixels in region 1 is 180, and the minimum gray value is 60. The initial contrast value is The above regional data are summarized in Table 2:

[0027] As shown in Table 2, the initial eigenvalues ​​of each region are normalized to their maximum and minimum values. Taking the edge strength value of region 1 as an example, the normalization calculation method is: (actual value - minimum value) / (maximum value - minimum value), that is: (10-6) / (10-6) = 1, the edge strength value of region 2 is normalized to (8-6) / (10-6) = 0.5, and the edge strength value of region 3 is 0; the texture fineness value and contrast value are normalized in a similar way, and finally the multi-feature standardized value matrix of each region is obtained, thereby generating multi-region standardized image eigenvalues.

[0028] S202: Calling multi-region standardized image feature values, collecting multi-region edge intensity values, texture fineness values, contrast values, gradient direction change values, spectrum balance values ​​and color distribution concentration, performing nonlinear cross integration operations, using the formula: ; Calculate and obtain regional characteristic response values, adjust the attribute response ratio within the region, and generate regional characteristic response distribution values; in, represents the characteristic response value of the bth region, represents the normalized edge strength value of the bth region, represents the normalized texture fineness value of the bth region, represents the normalized contrast value of the bth region, Indicates the gradient direction change value of the b-th region, represents the spectrum balance value of the bth area, represents the color distribution concentration of the bth area, Represents the joint summation of spectrum and color; Call the multi-region standardized image feature values ​​generated above and select the normalized edge intensity value , Normalized texture fineness value , normalized contrast value As the initial feature parameter. In addition, it is necessary to further measure and calculate the gradient direction change value of each area , spectrum balance value , Color distribution concentration For example, for region 1, through the statistical analysis of gradient direction, if the standard deviation of the pixel gradient direction in the region is measured to be 45°, then the gradient direction change value ; The spectrum balance value is quantitatively evaluated by the uniformity of the regional spectrum curve distribution. Assuming that the goodness of fit of the spectrum curve in region 1 is 0.85, the spectrum balance value ; The color distribution concentration is calculated by the regional RGB color variance. Assuming that the color distribution variance of region 1 is 0.2, the color distribution concentration is The above parameter data for area 1 are listed as follows: Normalized edge strength value , normalized texture fineness value , normalized contrast value , gradient direction change value , spectrum balance value , color distribution concentration ; Substitute the above data into the formula: ; The specific operation process is as follows: First calculate the numerator: ; Then calculate the denominator, which is the sum of the spectrum balance and color concentration of each area. Assuming that the data of all three areas are as shown in Table 2, we have: ; Therefore, the characteristic response value of region 1 is: ; Similarly, the characteristic response values ​​of regions 2 and 3 are calculated, and then the characteristic response values ​​of each region are obtained to form the regional characteristic response distribution value.

[0029] The benefit of the formula is that, by introducing the comprehensive quantification of gradient direction change value, spectrum balance and color concentration, the ability to distinguish regional feature responses is effectively improved.

[0030] S203: According to the regional feature response distribution values, the multi-region response value grid is mapped to the image space position, a regional weight distribution matrix is ​​established, and a regional feature response weight map is obtained.

[0031] The specific implementation process is to first define a spatial grid in the original image space, for example, divide the image into grid cells of size 10 pixels × 10 pixels, calculate the spatial position coordinates of each grid cell in the area based on the grid cells, and determine the label number of the area to which each grid cell belongs. . Assuming that the label corresponding to the grid cell where the spatial position (50,100) is located is region 1, and the feature response value of region 1 is 1.703 calculated above, the feature response value of the grid position (50,100) is set to 1.703. Similarly, the feature response values ​​of other spatial positions are mapped one by one. Subsequently, the feature response weight value in each grid cell is calculated by weighted averaging the feature response values ​​of adjacent spatial positions through a spatial smoothing function, such as a Gaussian kernel function (radius of 3 pixels), to generate a regional weight distribution matrix. Assuming that the feature response values ​​of the adjacent grid cells of the spatial position (50,100) are 1.703, 1.5, and 1.2, respectively, the weights (0.6, 0.3, and 0.1) are used for weighted averaging: , the regional feature response weight value of the position is obtained to be 1.592, and this method is used to perform weighted smooth mapping on the entire image, and finally obtain the regional feature response weight map.

[0032] See also Figure 4 , the steps for obtaining the image enhancement residual map are as follows: S301: Decomposing the input image according to the regional feature response weight map, extracting the edge-guided feature map, the texture map and the brightness map, performing regional feature response weight comparison, extracting high-frequency regional features, and generating multi-scale regional response feature values; First, the input image is scaled. Taking three scales as an example, Gaussian filtering is performed on the input image in turn, where the Gaussian filter kernel parameters are set to the standard deviation The scales are 0.5, 1.0, and 2.0 respectively, thus obtaining three image hierarchies of different scales. The above image hierarchies are used to extract edge-guided feature maps, texture maps, and brightness maps scale by scale. During the extraction process, taking scale 1.0 as an example, the Sobel operator is first called to perform convolution operation on the scale image to obtain the edge-guided feature map of the image, and then the local variance calculation is used to obtain the texture map. The brightness map is generated by the grayscale average value of the scale image and the grayscale difference of the neighborhood. The specific execution process is: the absolute value of the grayscale difference between each pixel and the grayscale value of the neighborhood (3×3 window) is calculated and averaged to obtain the brightness map. At this time, the brightness map value is obtained. Expressed as integer values, for example, the grayscale values ​​of the pixel area at a certain position are , then the brightness mapping value of the area Calculated as: ; Subsequently, for the above-mentioned images of different scales, the edge, texture, and brightness mapping data of the corresponding scales are extracted through the above steps and stored in arrays respectively. Then, the response values ​​of the edge-guided feature map, texture map, and brightness map are compared. Based on the local area (for example, a 5×5 neighborhood), the above-mentioned three eigenvalues ​​are normalized respectively, and the regional feature response weight comparison is performed pixel by pixel. By performing difference calculation on the above-mentioned three normalized eigenvalues, the threshold is set to 0.1 (the threshold is set based on the fact that when the difference in eigenvalues ​​exceeds 10% of the average value, it is a high-frequency area). If the edge-guided eigenvalue is greater than the threshold and is significantly higher than the texture and brightness features, the area is marked as a high-frequency area and the eigenvalue of the area is retained, otherwise it is ignored, thereby obtaining the high-frequency area feature data, and generating multi-scale regional response eigenvalues ​​in this way.

[0033] S302: Call the multi-scale regional response characteristic value, calculate the response difference value, construct the characteristic sensitivity coefficient, and adjust the residual response with the brightness map difference value, using the formula: ; Calculate and obtain the weight-adjusted residual structure; in, represents the weight-adjusted residual structure, represents the intensity difference of the edge-guided feature map at the cth scale, represents the structural difference of the texture map at the cth scale, Represents the grayscale distribution value of the brightness map at the cth scale, represents the characteristic sensitivity coefficient, Indicates the maximum value operation among all scales; For each scale, the edge-guided feature intensity difference is calculated separately and texture map structure difference ,in The specific calculation process is the absolute value of the difference between the edge features of adjacent pixels. For example, if the edge response values ​​of the position (x, y) and the neighborhood (x+1, y) are 0.7 and 0.4 respectively, then the calculation ; while texture mapping structure difference is the absolute value of the difference in texture mapping values ​​in adjacent areas. For example, if the texture values ​​at the same position (x, y) and the neighboring area (x+1, y) are 0.6 and 0.2 respectively, then .

[0034] Then use the brightness mapping value (See example in paragraph S301 for calculation) and the calculated response difference value to construct the characteristic sensitivity coefficient ,in The average value of the response difference in the local area is used as the sensitivity coefficient. For example, for a 5×5 area, if the average response difference in the area is 0.25, the characteristic sensitivity coefficient is set to , and further call the formula to calculate the residual response weight adjustment structure quantity : ; Combine the above sample data into the calculation, taking scale 1 as an example: ; By calculating all scales and taking the maximum value as the final weight to adjust the residual structure, for example, the calculation results of scales 1, 2, and 3 are 0.0070, 0.0095, and 0.0062 respectively, the final The maximum value is 0.0095.

[0035] S303: adjusting the residual structure amount according to the weight, fusing the residual signals of all scale maps, constructing the multi-scale structural relationship of the image elements, and generating an image enhancement residual map.

[0036] Bring in the above parameters The result value of the residual structure is adjusted by weight, and the residual signals of all scale images are fused scale by scale. The specific execution process includes: first, for each scale image, the residual structure is multiplied by the residual signal value of the original scale image pixel by pixel, for example, Figure 1 The original residual signal of a pixel is 0.4, then after fusion calculation it is ; Similar operations are performed scale by scale to obtain image data after residual adjustment at multiple scales. Finally, weighted fusion processing is performed pixel by pixel on the image data after residual adjustment at different scales. The weight of weighted fusion is determined according to the sensitivity coefficient of the original feature at different scales. The ratio is determined. For example, if the characteristic sensitivity coefficients corresponding to scales 1, 2, and 3 are 0.25, 0.30, and 0.20 respectively, then the scale Figure 1 , 2 The fusion weights of 3 are 33.3%, 40.0% and 26.7% respectively. The calculation of the residual map after pixel fusion at a certain position is:

[0037] This method is used to complete the weighted fusion of pixels at all positions, and finally construct a complete image enhancement residual map.

[0038]

[0039] As shown in Table 3, the data in the table show the calculation results of regional response characteristics and sensitivity coefficients at three scales, among which the maximum Used for subsequent scale fusion processing.

[0040] See also Figure 5 , the specific steps for obtaining the multi-scale enhanced fusion image are: S401: Based on the image enhancement residual map, call the scale map brightness mean and edge direction distribution map, calculate the brightness offset difference and the edge direction angle value, select the area that meets the edge consistency threshold condition, and obtain the edge consistency area distribution coefficient; First, call the brightness mean of the scalogram. In the specific execution process, assume that the scalogram is divided into multiple sub-regions. For each region, calculate the sum of the pixel brightness values ​​and divide it by the number of pixels in the region to obtain the regional brightness mean. For example, assume that there are 10 pixels in a certain scalar region, and their brightness values ​​are: , then by summing , and then divided by the total number of pixels 10, we get the average brightness of the area Similarly, the corresponding brightness mean data is calculated for other scale areas; then, the edge direction distribution map is called. In the specific implementation process, the directional gradient method is used to extract the image edge gradient vector pixel by pixel, and the gradient direction angle value at each pixel is calculated. For example, the Sobel operator is used to calculate the gradient amplitude in the horizontal and vertical directions point by point. , and then calculate for each pixel , if the horizontal gradient value of a pixel is 15 and the vertical gradient value is 10, then the edge direction angle at the pixel is calculated as ; After that, the brightness offset difference is calculated. The specific execution process is to call the brightness value in the residual map pixel by pixel and perform the difference operation with the above brightness mean. For example, if the brightness value of a pixel is 164, the brightness offset difference is calculated as ; Then calculate the correlation between the brightness offset difference and the edge direction angle. Specifically, compare the angle between the brightness difference direction between pixels and the edge direction, that is, calculate the absolute difference between the gradient direction of brightness change at each pixel and the angle between the edge direction. If the brightness change gradient direction at a certain pixel is , the edge direction angle is , then the angle difference is Next, the area screening action is performed for the calculated angle difference, that is, the pixel area whose angle difference is less than the preset edge consistency threshold is judged and screened. The specific edge consistency threshold setting is obtained through the previous image data statistics. It is assumed that the angle difference distribution of 1000 typical images is analyzed experimentally to determine the angle difference within the preset edge consistency threshold. The range belongs to the area with high consistency, and the threshold is selected as , then the area that meets the conditions is determined to be an edge consistent area; further, for the edge consistent area, the number of pixels that meet the threshold is counted and divided by the total number of pixels to calculate the edge consistent area distribution coefficient. Assuming that the number of pixels that meet the threshold condition is 680 and the total number of pixels involved in the calculation is 1000, the edge consistent area distribution coefficient is calculated as , thus obtaining the edge consistent area distribution coefficient.

[0041] S402: Extract the texture contrast value of the matching area according to the edge consistent area distribution coefficient, and construct the brightness difference and texture contrast difference ratio using the formula: ; Calculate the ratio of overall brightness deviation to texture distinctness, reorder the regional brightness according to texture features, and obtain a brightness callback reference sequence; in, Represents the ratio of overall brightness deviation to texture sharpness. Indicates the number of pixels involved in the calculation in the edge consistent area. Represents the brightness value of the dth pixel in the edge consistent area, represents the mean brightness of the pixel in the scale map, Indicates the texture contrast value corresponding to the d-th pixel in the texture contrast map, Indicates the maximum value of the texture contrast value corresponding to all participating pixels in the current area. Represents the brightness difference between the dth pixel and the adjacent rth pixel; Substitute the above parameters into the result value of the edge consistent area distribution coefficient , and then call the texture contrast value of the matching area. In the specific execution process, the texture contrast value is calculated by calling the image gray level co-occurrence matrix (GLCM) to quantitatively analyze the gray level difference of the area near each pixel. Taking a certain pixel as an example, the pixel gray values ​​in its local window (for example, a 3×3 neighborhood window) are , the texture contrast value of this pixel is calculated according to GLCM is 7.2. Similarly, the texture contrast is calculated point by point for all pixels. Then, the ratio of the brightness difference to the texture contrast is calculated. The participating parameters in the formula are expressed as: Indicates the total number of pixels in the edge consistent area, which is set to 680 in this example. Indicates the brightness value of the dth pixel. For example, the brightness value of the pixel is 164. is the mean brightness (158.4), then , and then with its texture contrast value Sum, get , after calculating and accumulating the 680 pixels separately, assume that the accumulated sum is 8500, then calculate the maximum value of the texture contrast in the participating pixel area, for example, assume it is 12.0, and calculate the brightness difference value between the pixels at the same time For example, if the brightness of two adjacent pixels is 164 and 160 respectively, then the brightness difference is , accumulate all adjacent pixel differences to calculate the sum, assuming it is 3000, and substitute it into the formula for calculation: ; The result shows that the ratio of overall brightness deviation to texture vividness is 0.236. The ratios of each region are reordered from large to small to obtain the brightness callback reference sequence of the corresponding region.

[0042] S403: Based on the brightness callback reference sequence, call the brightness information of the matching area, perform brightness mapping and segment adjustment, integrate texture features and edge direction features, and obtain a multi-scale enhanced fusion image.

[0043] Substitute the above parameter brightness callback reference sequence result, call the matching area brightness information, perform brightness mapping and segmented adjustment on the brightness value pixel by pixel. In the specific execution process, set the mapping function to linear transformation by region. Assuming that the original brightness range of a region is [150,170], the brightness range after adjustment is mapped to [155,175]. When the original brightness of a pixel is 164, its corresponding brightness after adjustment is: ; The adjusted brightness is integrated with the original texture features and edge direction features, and the pixel-by-pixel fusion operation is performed through the multi-scale fusion technology. In the specific fusion process, it is assumed that a fusion weight is assigned to each scale image, and the weight depends on the comprehensive index of regional texture clarity and edge direction consistency. For example, the regional texture clarity index is 0.72, the edge direction consistency coefficient is 0.68, and the fusion weight is set to , multiply the weight with the brightness adjusted image of the corresponding scale, then add and fuse them scale by scale to finally form a multi-scale enhanced fused image.

[0044] See also Figure 6 , the specific steps for obtaining the regional enhanced image result map are: S501: according to the multi-scale enhanced fusion image, call the edge response difference map of the original image and the enhanced image, obtain the regional edge response strength, compare the gradient response value of the difference map at the regional edge position with the set edge enhancement discrimination threshold, and determine whether the region belongs to the high-frequency change region, and generate the high-frequency response interval value; The specific execution process is as follows: first, call the edge response values ​​of the corresponding positions of the original image and the enhanced image, subtract the edge response values ​​of the corresponding positions in the enhanced image from the edge response values ​​of the corresponding positions in the original image, and obtain the edge response difference map corresponding to the positions of each area. Taking an actual scene as an example, randomly select a position in the monitored image. Assuming that the edge response value of the original image is , the enhanced image edge response value is , then the difference is , the value of the difference map at this position is Next, the gradient response value at the edge of the region is calculated using the numerical difference method for the difference map. For example, the adjacent positions Take the response values ​​of the difference graph and set them as , , , At this time, the gradient response value of this position is calculated by finite difference: Next, the gradient response value is specifically compared with the set edge enhancement discrimination threshold. The edge enhancement discrimination threshold is set based on the experimental analysis of edge characteristics. For example, in the experiment, the edge enhancement discrimination threshold is taken as , current regional location Gradient response value , thus determining that the position belongs to the high-frequency change area, and the corresponding difference map value Record the high frequency response interval value. Similarly, the other positions in the image are traversed and executed in the same way to obtain the high frequency response interval value set of the entire image.

[0045] S502: Call the high-frequency response interval value, combine the texture consistency offset value and the brightness change rate, and use the formula for all areas in the high-frequency response interval: ; Calculate the enhancement amplitude index, filter the area greater than the edge enhancement discrimination threshold, and obtain the edge enhancement effective coefficient distribution; in, Represents the enhancement amplitude index of the coordinate (e, s) position, The texture consistency offset value representing the position, represents the rate of change of brightness, represents the response value of the edge response difference map, represents the enhancement value of the x-th layer fusion image at the position under the differentiated scale, and v represents the number of scales involved in the fusion image; Call the high frequency response interval value obtained above and the texture consistency offset value and brightness change rate , respectively, are obtained by analyzing the local features of the image, where the texture consistency offset value uses the local standard deviation algorithm to analyze the difference between the local pixel and the central pixel, such as the position The local standard deviation of , then the texture consistency offset value is ; The brightness change rate is calculated by the grayscale difference of the image pixels before and after enhancement. For example, the grayscale value of the original image is , the gray value of the enhanced image is , the brightness change rate is . At the same time, obtain the fusion enhancement value under the differentiated scale , assuming that the three-layer fusion scale is currently used, the specific fusion enhancement value is expressed as follows using an actual example:

[0046] As shown in Table 4, the fusion enhancement values ​​of this area are 1.5, 1.2, and 0.8 respectively. Corresponding position for , substitute the above values ​​into the formula: ; After the above process is performed for all high-frequency response interval positions, the enhancement amplitude index of each position is obtained. Next, the enhancement amplitude index of each position is compared and screened one by one, and the aforementioned edge enhancement discrimination threshold is continued. To make a practical comparison, the location The enhancement index , then this position does not belong to the effective position of edge enhancement. If the enhancement amplitude index of another position is , then this position is regarded as the effective area of ​​edge enhancement. After the above comparison and screening one by one, a complete distribution map of effective coefficients of edge enhancement is finally obtained.

[0047] S503: Calling the edge enhancement effective coefficient distribution, identifying the area below the edge enhancement discrimination threshold, restoring the image value to the original image value, and maintaining the enhancement processing for the remaining areas to obtain the regional enhanced image result map.

[0048] Call the edge enhancement effective coefficient distribution map obtained above to judge and analyze all regional positions one by one, for example, The corresponding enhancement index is , judged to be below the edge enhancement discrimination threshold , then restore the image value at that position to the original image value, that is If another location The enhancement index is , then the position keeps the value after enhancement processing After performing the above operations on all positions in sequence, a complete region enhanced image result map is obtained.

[0049] An image enhancement system based on deep learning, the image enhancement system based on deep learning is used to execute the above-mentioned image enhancement method based on deep learning, and the system includes: The image frequency division module obtains the input image pixel matrix, extracts the brightness difference, gradient direction change value and local texture density value, makes judgments based on the brightness difference and density product and the gradient direction change value, divides the frequency area and marks the image block type, and generates the image frequency partition label; The regional feature mapping module calls the image frequency partition label, extracts the regional edge intensity value, texture fineness value and contrast value, normalizes them and weightedly superimposes them to the corresponding image block position to generate a regional feature response weight map; The scale residual fusion module calls the regional feature response weight map, obtains the image block content and extracts the multi-scale edge-guided map, texture map and brightness map. It performs residual calculation and weighted fusion on the three maps according to the image block weight value. The fusion value is superimposed according to the image block position to reconstruct the entire image and generate an image enhancement residual map. The edge brightness callback module calls the image enhancement residual map, collects the brightness mean, edge direction distribution map and texture contrast map, determines the regional matching relationship based on the edge direction consistency coefficient and performs brightness linear callback to generate a multi-scale enhanced fusion image; The high-frequency detail restoration module calls the multi-scale enhanced fusion image, extracts the edge response difference map between it and the original image, and makes a combined judgment based on the texture consistency offset value and the brightness change rate. It determines that the unsatisfactory area is restored to the original image value and generates a regional enhanced image result map.

[0050] The above are only preferred embodiments of the present invention and are not intended to limit the present invention in other forms. Any technician familiar with the profession may use the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. An image enhancement method based on deep learning, characterized in that: The following steps are involved: S1: Get the pixel matrix of the input image, determine the regional frequency attribute according to the brightness difference, gradient direction change value and local texture density, divide it into high-frequency area and low-frequency area according to the edge response threshold, and obtain the image frequency partition label; S2: extracting edge intensity values, texture fineness values ​​and contrast values ​​of multiple regions based on the image frequency partition labels, performing normalized weighted superposition operations, assigning the results to corresponding regions, and obtaining a regional feature response weight map; S3: According to the regional feature response weight map, the input image is scaled and decomposed, a multi-scale edge-guided feature map, a texture map and a brightness map are extracted, the multi-map residual response value is adjusted according to the weight, the high-frequency regional details are enhanced and the low-frequency regional redundancy is suppressed, all scale maps are fused and a mapping structure is constructed to obtain an image enhancement residual map; S4: Based on the image enhancement residual map, call the scale map brightness mean, edge direction distribution map and texture contrast map, match the area according to the edge consistency coefficient and perform brightness callback to obtain a multi-scale enhanced fusion image.

2. The image enhancement method based on deep learning according to claim 1, characterized in that: The image frequency partition labels are specifically high-frequency areas and low-frequency areas. The regional feature response weight map includes multi-region edge intensity values, texture fineness values, and contrast values. The image enhancement residual map is specifically a multi-scale edge-guided feature map, a texture map, and a brightness map after scale decomposition. The multi-scale enhanced fusion image includes a scale map brightness mean, an edge direction distribution map, and a texture contrast map.

3. The image enhancement method based on deep learning according to claim 2, characterized in that: The steps of obtaining the image frequency partition label are specifically as follows: S101: Based on the input image pixel matrix, the grayscale difference is calculated and compared with the brightness difference threshold, the number of pixels with significant brightness difference is counted and normalized, and the brightness difference density value is obtained; S102: calling the brightness difference density value, combining the multi-pixel gradient direction change difference, the neighborhood gradient angle difference, and the local change trend intensity, using the formula: ； The regional gradient response value is obtained by calculation and superimposed with the distribution of brightness difference density to obtain the frequency mutation factor value; in, Represents the frequency mutation factor value, Represents the gradient direction change difference of the a-th pixel pair, Indicates the trend strength of the local gradient direction change at the point, is the point gradient change density value, is the brightness difference density value of the point neighborhood, and N represents the number of pixel pairs involved in the statistics; S103: comparing the frequency mutation factor value with the texture density distribution and the edge response threshold, determining the regional frequency attribute, and generating an image frequency partition label.

4. The image enhancement method based on deep learning according to claim 3, characterized in that: The steps for obtaining the regional feature response weight map are specifically as follows: S201: extracting edge intensity values, texture fineness values, and contrast values ​​of multiple regions based on the image frequency partition labels, normalizing them according to the maximum and minimum values, and generating multi-region standardized image feature values; S202: calling the multi-region standardized image feature values, collecting multi-region edge intensity values, texture fineness values, contrast values, gradient direction change values, spectrum balance values ​​and color distribution concentration, performing nonlinear cross integration operations, using the formula: ； Calculate and obtain regional characteristic response values, adjust the attribute response ratio within the region, and generate regional characteristic response distribution values; in, represents the characteristic response value of the bth region, represents the normalized edge strength value of the bth region, represents the normalized texture fineness value of the bth region, represents the normalized contrast value of the bth region, Indicates the gradient direction change value of the b-th region, represents the spectrum balance value of the bth area, represents the color distribution concentration of the bth area, Represents the joint summation of spectrum and color; S203: According to the regional feature response distribution values, a multi-region response value grid is mapped to an image space position, a regional weight distribution matrix is ​​established, and a regional feature response weight map is obtained.

5. The image enhancement method based on deep learning according to claim 4, characterized in that: The steps of obtaining the image enhancement residual map are specifically as follows: S301: Decomposing the input image on a scale according to the regional feature response weight map, extracting the edge-guided feature map, the texture map and the brightness map, performing regional feature response weight comparison, extracting high-frequency regional features, and generating multi-scale regional response feature values; S302: Call the multi-scale regional response characteristic value, calculate the response difference value, construct the characteristic sensitivity coefficient, and adjust the residual response with the brightness map difference value, using the formula: ； Calculate and obtain the weight-adjusted residual structure; in, represents the weight-adjusted residual structure, represents the intensity difference of the edge-guided feature map at the cth scale, represents the structural difference of the texture map at the cth scale, Represents the grayscale distribution value of the brightness map at the cth scale, represents the characteristic sensitivity coefficient, Indicates the maximum value operation among all scales; S303: adjusting the residual structure amount according to the weight, fusing the residual signals of all scale maps, constructing a multi-scale structural relationship of image elements, and generating an image enhancement residual map.

6. The image enhancement method based on deep learning according to claim 5, characterized in that: The steps of acquiring the multi-scale enhanced fusion image are specifically as follows: S401: Based on the image enhancement residual map, call the scale map brightness mean and edge direction distribution map, calculate the brightness offset difference and the edge direction angle value, select the area that meets the edge consistency threshold condition, and obtain the edge consistency area distribution coefficient; S402: Extract the texture contrast value of the matching area according to the edge consistent area distribution coefficient, and construct the brightness difference and texture contrast difference ratio using the formula: ； Calculate the ratio of overall brightness deviation to texture distinctness, reorder the regional brightness according to texture features, and obtain a brightness callback reference sequence; in, Represents the ratio of overall brightness deviation to texture sharpness. Indicates the number of pixels involved in the calculation in the edge consistent area. Represents the brightness value of the dth pixel in the edge consistent area, Represents the mean brightness of the pixel in the scale map, Indicates the texture contrast value corresponding to the d-th pixel in the texture contrast map, Indicates the maximum value of the texture contrast value corresponding to all participating pixels in the current area. Represents the brightness difference between the dth pixel and the adjacent rth pixel; S403: Based on the brightness callback reference sequence, call the brightness information of the matching area, perform brightness mapping and segmented adjustment, integrate texture features and edge direction features, and obtain a multi-scale enhanced fused image.

7. The image enhancement method based on deep learning according to claim 6, characterized in that: The method further comprises: S5: According to the multi-scale enhanced fusion image, the edge response difference map, texture consistency offset value and brightness change rate of the original image and the enhanced image are called, and the enhancement amplitude judgment operation is performed on all high-frequency areas in the image, and the area with a response amplitude lower than the edge enhancement judgment threshold is eliminated, and the original image value is restored to it, and the regional enhanced image result map is output; The region enhanced image result map is specifically an image after the high-frequency region enhancement amplitude is determined.

8. The image enhancement method based on deep learning according to claim 7, characterized in that: The steps for obtaining the region enhanced image result map are specifically as follows: S501: According to the multi-scale enhanced fusion image, call the edge response difference map of the original image and the enhanced image to obtain the regional edge response strength, compare the gradient response value of the difference map at the regional edge position with the set edge enhancement discrimination threshold, and determine whether the region belongs to the high-frequency change region, and generate a high-frequency response interval value; S502: calling the high frequency response interval value, combining the texture consistency offset value and the brightness change rate, and using the formula for all areas in the high frequency response interval: ； Calculate the enhancement amplitude index, filter the area greater than the edge enhancement discrimination threshold, and obtain the edge enhancement effective coefficient distribution; in, Represents the enhancement amplitude index of the coordinate (e, s) position, The texture consistency offset value representing the position, represents the rate of change of brightness, represents the response value of the edge response difference map, represents the enhancement value of the x-th layer fusion image at the position under the differentiated scale, and v represents the number of scales involved in the fusion image; S503: calling the edge enhancement effective coefficient distribution, identifying the area below the edge enhancement discrimination threshold, restoring the image value to the original image value, and maintaining the enhancement process for the remaining areas to obtain the regional enhanced image result map.

9. The image enhancement system based on deep learning is characterized by: According to the deep learning-based image enhancement method according to any one of claims 1 to 8, the system comprises: The image frequency division module obtains the input image pixel matrix, extracts the brightness difference, gradient direction change value and local texture density value, makes judgments based on the brightness difference and density product and the gradient direction change value, divides the frequency area and marks the image block type, and generates the image frequency partition label; The regional feature mapping module calls the image frequency partition label, extracts the regional edge intensity value, texture fineness value and contrast value, normalizes them and weightedly superimposes and assigns them to the corresponding image block position to generate a regional feature response weight map; The scale residual fusion module calls the regional feature response weight map, obtains the image block content and extracts the multi-scale edge-guided map, texture map and brightness map, performs residual calculation and weighted fusion on the three maps according to the image block weight value, and reconstructs the whole image by superimposing the fusion value according to the image block position to generate an image enhancement residual map; The edge brightness callback module calls the image enhancement residual map, collects the brightness mean, edge direction distribution map and texture contrast map, determines the regional matching relationship according to the edge direction consistency coefficient and performs brightness linear callback to generate a multi-scale enhanced fusion image; The high-frequency detail restoration module calls the multi-scale enhanced fusion image, extracts the edge response difference map between it and the original image, and makes a combined judgment based on the texture consistency offset value and the brightness change rate, judges that the unsatisfactory area is restored to the original image value, and generates a regional enhanced image result map.

Images