Image enhancement method based on fusion parameter filter

Abstract

The invention belongs to the technical field of image enhancement, and particularly relates to an image enhancement method based on a fusion parameter filter. The method comprises the following steps: adjusting weights of a step filter, an elliptic filter and a polynomial filter according to proportions of three types of features including gray features, texture features and structural features in a scene image; and inputting the extracted scene image into a fusion parameter filter to obtain a filtered image, wherein the fusion parameter filter is a filter obtained by carrying out weighted summation on the three filters by using the adjusted weights of the three filters. According to the invention, the weight of each filter in the fusion parameter filter is adaptively adjusted according to the feature change condition in the scene image, so that the fusion parameter filter can adaptively change according to the change of the scene in the scene image, and a better image enhancement effect can be obtained no matter what scene is.

Description

Technical Field

[0001] The invention belongs to the technical field of image enhancement, and in particular relates to an image enhancement method based on a fusion parameter filter. Background Art

[0002] During the image acquisition process, we often encounter poor environmental visibility conditions. At this time, the image quality obtained is poor, especially in fog, rain, snow, dust and other weather conditions. The images obtained are even worse, mainly manifested in low contrast, color distortion, unclear image details, etc., which not only affects the visual effect of the observed image, but also limits the normal operation of the outdoor vision system.

[0003] In order to overcome this difficulty, it is necessary to filter and enhance the acquired degraded image. Traditional image enhancement methods mainly use a single filtering algorithm for filtering and enhancement, but a filtering algorithm often focuses on a certain point, simply achieving local enhancement, and its application scenario is often single, that is, the filtering algorithm is only applicable to image enhancement of images in a certain application scenario. If you change the scenario, the filtering algorithm will not be applicable, resulting in poor filtering and enhancement effects for images in other application scenarios. Summary of the invention

[0004] The purpose of the present invention is to provide an image enhancement method based on a fusion parameter filter, so as to solve the problem that the image enhancement effect of some scenes is poor due to the single filtering algorithm used in the prior art being not suitable for all application scenarios.

[0005] In order to solve the above technical problems, the present invention provides an image enhancement method based on a fusion parameter filter, the method comprising:

[0006] Acquire a scene image, extract the proportions of its grayscale features, texture features and structural features, and adjust the weights of the step filter, the elliptic filter and the polynomial filter according to the proportions of the three types of features so that: when the difference between the proportions of any two types of features among the three types of features is less than a set difference threshold, the weights of the three filters are equal; otherwise, when the proportions of the grayscale features, texture features and structural features are respectively the largest, the weights of the step filter, the elliptic filter and the polynomial filter are correspondingly the largest;

[0007] The scene image is input into a fusion parameter filter to obtain a filtered scene image; the fusion parameter filter is a filter obtained by weighted summing the three filters using the weights adjusted by the three filters.

[0008] Furthermore, in other cases, when the proportions of grayscale features, texture features and structural features are respectively the smallest, the weights of the step filter, elliptical filter and polynomial filter are correspondingly minimized; or in other cases, the weights of the filters corresponding to the two types of features with smaller proportions are made equal; wherein the grayscale features, texture features and structural features correspond to the step filter, elliptical filter and polynomial filter respectively.

[0009] Furthermore, the method for determining the proportion of the three types of features is as follows: normalize the vector representations corresponding to the three types of features respectively, convert the normalized vectors into numerical values ​​to obtain the numerical values ​​corresponding to the three types of features respectively; calculate the proportion of the numerical value corresponding to each type of feature to the sum of the corresponding numerical values ​​of the three types of features, and the proportion is the proportion corresponding to each type of feature.

[0010] Furthermore, the vector obtained by calculating the grayscale histogram of the scene image is the vector representation of the grayscale feature, the vector obtained by calculating the grayscale co-occurrence matrix of the scene image is the vector representation of the texture feature, and the vector obtained by calculating the gradient direction histogram of the scene image is the vector representation of the structural feature.

[0011] Furthermore, the method also includes: performing image downsampling of different scales on the filtered scene image, and extracting high-frequency components of the obtained downsampled images to obtain high-frequency components of different scales; extracting feature information of different scales on the filtered scene image through an encoder, and then using a decoder to map the feature information of different scales output by the encoder back to the original data space, and the output of each layer structure in the encoder is a result obtained by performing jump connection on the corresponding scale feature information output by the encoder, and fusing it with the high-frequency components of the same scale, and then performing feature extraction on the fused result.

[0012] Furthermore, the method of adjusting the weights of the three filters includes: if the proportion of grayscale features is greater than or equal to 0.5, the weight of the step filter is 0.5, and the weights of the elliptic filter and the polynomial filter are both 0.25; if the proportion of texture features is greater than or equal to 0.5, the weight of the elliptic filter is 0.5, and the weights of the step filter and the polynomial filter are both 0.25; if the proportion of structural features is greater than or equal to 0.5, the weight of the polynomial filter is 0.5, and the weights of the step filter and the elliptic filter are both 0.25.

[0013] Furthermore, octave convolution is used to extract high-frequency components.

[0014] Furthermore, the acquired scene image is an image that has been subjected to denoising processing.

[0015] Furthermore, the expression of the polynomial filter is s3(x,y)=Img*(x+y+γ)p Or s3(x,y)=(x+y+Img+γ) p , where s3 represents a polynomial filter, Img represents the intensity of the image pixel at the (x, y) position, γ represents an independent scalar, and p = 3.

[0016] Furthermore, the method also includes: calculating the size of the evaluation index of the final image, the evaluation index including at least one of the structural similarity, contrast sensitivity function, peak signal-to-noise ratio and minimum average error; evaluating the image enhancement effect according to the evaluation index, and if the effect is not good, further adjusting the weights corresponding to the three filters.

[0017] The present invention is a pioneering invention, and its beneficial effects are: considering that the step filter has a good effect on images in high-contrast plane areas, the elliptical filter can adjust images with elliptical feature areas, and the polynomial filter can adjust the refinement of the entire image to achieve a local enhancement effect, these three filters can meet most scene requirements. The present invention proposes a new fusion parameter filter, which is used to fuse the filtering results of the above three filters. Moreover, when fusing the filtering results of these three filters, the weights of these three filters are adjusted according to the proportions of grayscale features, texture features and structural features in the scene image. When the proportions of the grayscale features, texture features and structural features are not much different, the weights of the three filters can be directly set equal. When the proportion of grayscale features is large, the weight of the step filter is set to be larger, when the proportion of texture features is large, the weight of the elliptical filter is set to be larger, and when the proportion of structural features is large, the weight of the polynomial filter is set to be larger. The weights of each filter in the fusion parameter filter are adaptively adjusted according to the feature changes in the scene image, so that the fusion parameter filter can adapt to changes according to the changes in the scene in the scene image, and better image enhancement effects can be obtained regardless of the scene. It is particularly suitable for image degradation scenes under fog, rain, snow and dust weather conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a flow chart of an image enhancement method based on fusion parameter filters of the present invention;

[0019] Figure 2 It is a structural diagram of the encoder, decoder and high-frequency information extraction of the present invention;

[0020] Figure 3 It is the principle diagram of the step filter used in the present invention;

[0021] Figure 4 It is a schematic diagram of the elliptical filter used in the present invention;

[0022] Figure 5It is a specific flow chart of the image enhancement method based on fusion parameter filter of the present invention. DETAILED DESCRIPTION

[0023] The core idea of ​​the present invention is to adaptively adjust the weights of each filter in the fusion parameter filter according to the feature changes in the scene image. Specifically, when the difference between the proportions of any two of the three types of features of the scene image, namely, the grayscale features, texture features and structural features, is less than the set difference threshold, the weights of the three filters are equal. Otherwise, when the proportions of the grayscale features, texture features and structural features are the largest, the weights of the step filter, elliptic filter and polynomial filter are maximized accordingly, so that the fusion parameter filter obtained by weighted summation of the three filters using the adjusted weights of the three filters can be adapted to different scenes, and better image enhancement effects can be obtained for different scenes. In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0024] like Figure 1 As shown, the specific implementation process of the image enhancement method based on fusion parameter filter of the present invention is as follows:

[0025] Step 1: Load the natural scene image I(x).

[0026] The loaded natural scene image can be acquired by shooting on-site with a camera mounted on a tripod, and in order to ensure a better image enhancement effect, the loaded scene image can be denoised here.

[0027] Step 2: Extract grayscale features, texture features, and structural features from the scene image obtained in step 1, determine the weights of the three features, and adjust the weights of the three filters, namely, the step filter, the elliptical filter, and the polynomial filter, according to the weights of the three features to adjust the filtering strategy.

[0028] 1) Determine the weight of the three types of features.

[0029] Among them, the vector representation of grayscale features can be calculated by grayscale histogram (GH), the vector representation of texture features can be calculated by gray level co-occurrence matrix (GLCM), and the vector representation of structural features can be calculated by gradient directional histogram (HOG).

[0030] Grayscale histogram is a statistical representation of the number of pixels with different grayscale values, using vector I vec Representation, that is, the vector representation of grayscale features. Vector I vec The i-th element I in vec (i) is:

[0031] I vec (i) = N(i)

[0032] Where N(i) represents the number of pixels with grayscale value i in the scene image, and the value of i is [0, 255], with a total of 256 grayscales.

[0033] The gray-level co-occurrence matrix converts the original m-row and n-column image into a 256*256 matrix I by counting the relative position relationship between pixels. glcm , record matrix I glcm The diagonal elements of get the vector I expressing the texture features g , which is the vector representation of texture features.

[0034] The structural features are extracted through the gradient direction histogram. The specific process is to first calculate the gradient of each pixel to capture the contour information, then divide the image into 6*6 cells, and count the gradient vector I of each cell. cell , every 4 cells form a block, and the block vector I is obtained block Finally, all block vectors are combined to obtain the HOG vector I hog , which is the vector representation of the structural features.

[0035] In order to make the grayscale, texture, and structural features comparable, they are first normalized:

[0036]

[0037] For the normalized vectors of the above three types of features, the three vectors are converted into numerical values ​​(specifically, the three vectors can be converted into three numerical values ​​by finding the modulus of the vector. Of course, other processing methods can also be used to convert the three vectors into three numerical values ​​to facilitate the calculation of the proportion), which is X' GH , X' GLCM , X' HOG Then, we calculate the weight of the three types of features (σ GH ,σ GLCM ,σ HOG ), calculated as follows:

[0038]

[0039] 2) Adjust the weights of the three filters according to the proportion of the three types of features.

[0040] The specific adjustment strategy is: when the difference between the proportions of any two of the three types of features is less than the set difference threshold, it indicates that the proportions of the three types of features in the scene image are relatively close, so that the weights of the three filters can be directly equal; otherwise, it means that the proportions of the three types of features are not very close. When the grayscale feature is relatively large, the weight of the step filter is relatively large, when the texture feature is relatively large, the weight of the elliptical filter is relatively large, and when the structural feature is relatively large, the weight of the polynomial filter is relatively large. Finally, when the proportions of grayscale features, texture features and structural features are the largest, the weights of the step filter, elliptical filter and polynomial filter are maximized.

[0041] Moreover, in the above-mentioned otherwise situation, the following two methods can be adopted. Method 1: When the proportions of grayscale features, texture features and structural features are respectively minimized, the weights of the step filter, elliptical filter and polynomial filter can be minimized accordingly. In other words, the processing method of method 1 is that the size relationship of the proportions of grayscale features, texture features and structural features is the same as the size relationship of the weights of the step filter, elliptical filter and polynomial filter. Method 2: Directly make the weights of the filters corresponding to the two types of features with smaller proportions equal. The corresponding relationship here is that the grayscale features, texture features and structural features correspond to the step filter, elliptical filter and polynomial filter respectively. In other words, the processing method of method 2 is to take into account that the two types of features with smaller proportions are not so large, and at this time, the weights of the filters corresponding to these two types of features are directly treated equally.

[0042] For example, when the proportion of grayscale features is greater than or equal to 0.5, the weight of the step filter is 0.5 and the weights of the elliptic filter and the polynomial filter are equal. Of course, the weights of the elliptic filter and the polynomial filter can also be determined according to the proportions of texture features and structural features, respectively. When the proportion of texture features is greater than the proportion of structural features, the weight of the elliptic filter is appropriately greater than the weight of the polynomial filter; if the proportion of texture features is greater than or equal to 0.5, the weight of the elliptic filter is 0.5 and the weights of the step filter and the polynomial filter are equal. Of course, the step filter and the polynomial filter can also be determined according to the proportions of texture features and structural features, respectively. The weight of the polynomial filter is determined according to the proportion of grayscale features and structural features respectively. When the proportion of grayscale features is greater than the proportion of structural features, the weight of the step filter is appropriately greater than the weight of the polynomial filter; if the proportion of structural features is greater than or equal to 0.5, the weight of the polynomial filter is 0.5 and the weights of the step filter and the elliptic filter are equal. Of course, the weights of the step filter and the elliptic filter can also be determined according to the proportion of grayscale features and texture features respectively. When the proportion of grayscale features is greater than the proportion of texture features, the weight of the step filter is appropriately greater than the weight of the elliptic filter.

[0043] For example, the weights of grayscale features, texture features, and structural features are 0.7, 0.2, and 0.1, respectively. Then the corresponding weights of the step filter, elliptical filter, and polynomial filter are 0.7, 0.2, and 0.1, respectively.

[0044] Step three, input the scene image obtained in step one into the constructed fusion parameter filter to achieve filtering of the scene image; the fusion parameter filter is a filter obtained by weighted summing the three filters using the weights of the three filters obtained in step two.

[0045] 1) The following is an introduction to the three types of filters.

[0046] Step filters are used to adjust images in high-contrast flat areas. Figure 3 As shown. The step filter consists of three parallel lines. The center line defines the position and direction of the filter in the form of y=kx+c, where k is the slope and c is the intercept. o1 and o2 are two additional offset parameters, so that each adjustment map consists of four different regions of the heat map s(x,y). In the 100% area, all pixels are multiplied by the scaling parameter s g ; In the 100%-50% area, the applied scaling factor is from s g to g / 2 linearly decreases; in the 50%-0% area, the scaling value further decreases linearly until it reaches the 0% area, in which the pixels are not adjusted. The mathematical description of the step filter s1(x,y) is as follows:

[0047]

[0048] Where a(x,y)=s g min{(1+l(x,y) / d1) / 2,1},b(x,y)=s g max{(1+l(x,y) / d2) / 2,1},s g is the scale scaling parameter, which is a parameter to prevent the calculated value from being too large and causing image enhancement distortion. l(x,y)=y-(kx+c) is a function of the position of the point (x,y) relative to the center line. d1=o1*cosα, d2=o2*cosα, α=arctank, g inv is a binary indicator variable, parameter g inv The top and bottom lines are allowed to be reversed, and the reversal determines the position of the 100% zoom region relative to the center line. To achieve learnable reversals, we predict a binary indicator parameter g inv =(sgn(g inv )+1) / 2, where sgn represents the sign function, g inv is a real-valued prediction parameter, and its binary value range is ginv ∈{0,1}. g To explain further: For example, in the 100% area, all pixels are multiplied by the scaling parameter s g After calculation, most of the pixel values ​​in this area exceed 255, and the image is displayed as a grayscale value of 255.

[0049] The elliptical filter is used to adjust the image with elliptical feature areas. The principle is as follows Figure 4 The image of the high contrast flat area is defined by an ellipse whose parameters include the center (x0, y0), the major semi-axis a, the minor semi-axis b and the rotation angle θ. The scaling factor s is assumed at the center of the ellipse e is the largest, and then decreases linearly until it reaches the boundary; pixels outside the ellipse are not adjusted. Mathematically, the elliptical filter is defined as follows:

[0050] s2(x,y)=s e min{1-(A 2 +B 2 ),0}

[0051] In the formula, A 2 =[(xh)cosθ+(yk)sinθ] 2 / a 2 , B 2 =[(xh)sinθ+(yk)cosθ] 2 / b 2 , (h, k) represents the center position of the elliptical filter. The elliptical filter is usually used to enhance and highlight objects or specific areas of interest in the image.

[0052] The polynomial filter is based on the inherent information of the image and defines the corresponding geometric shape to achieve fine adjustment of the entire image and ultimately achieve local enhancement. Two forms of polynomial filters are defined here, namely Img*(x+y+γ) p and (x+y+Img+γ) p , where Img represents the intensity of the image pixel at the (x, y) position and γ is an independent scalar. It is empirically found that a cubic polynomial, i.e., p = 3, can provide both rich expressiveness and a limited set of parameters for image enhancement.

[0053] Next, substitute p=3 into two filters respectively. The first one is Img'(x,y)=Img*(x+y+γ) 3 , expanded as follows:

[0054] Img'(x,y)=Img*(x+y+γ) 3

[0055] =Img*(Ax 3 +Bx 2 y+Cx 2 +D+Exy 2 +Fxy+Gx+Hy 3 +Iy 2 +Jy)

[0056] In the formula, {A, B, C, ..., J} are 10 polynomial coefficients, representing the filter parameters. This filter can map an image with intensity Img to a new adjusted intensity Img'. Another filter is composed of 20 coefficients {A, B, C, ..., T}, expressed as Img'(x, y) = (x + y + Img + γ) 3 , expanded as follows:

[0057] Img'(x,y)=(x+y+Img+γ) 3

[0058] =Ax 3 +Bx 2 y+Cx 2 Img+Dx 2 +Exy 2 +FxyImg+Gxy+HxImg 2 +IxImg+Jx+Ky 3 +Ly 2 Img+My 2 +NyImg 2 +OyImg+Py+QImg 3 +RImg 2 +SImg+T

[0059] The polynomial filter takes into account spatial and intensity information while limiting the complexity of the mapping to a regularized form, achieving local pixel enhancement while ensuring that the transformation is locally smooth. Therefore, the final expression of the polynomial filter is s3(x,y)=Img'(x,y).

[0060] 2) The above three filters are fused to obtain a fusion parameter filter, and the final fusion parameter filter is s L (x, y). Among them, the fusion result of the step filter s1(x, y) and the elliptic filter s2(x, y) is as follows:

[0061] s12(x,y)=α*s1(x,y)+(1-α)*s2(x,y)

[0062] Where α is used to balance the information loss between the step filter s1(x, y) and the elliptic filter s2(x, y), and the value range is α∈(0,1). Furthermore, the fusion result of s12(x, y) and the polynomial filter is as follows:

[0063] s123(x,y)=β*s12(x,y)+(1-β)*s3(x,y)

[0064] Where β is used to balance the information loss between s12(x, y) and the polynomial filter s3(x, y), and its value range is β∈(0,1). s123(x, y) is the output of the fusion parameter filter, denoted as Img 1 =s123(x,y).

[0065] That is, the expression of the final fusion parameter filter is:

[0066] s L (x,y)=s123(x,y)=β*α*s1(x,y)+β*(1-α)*s2(x,y)+(1-β)*s3(x,y)

[0067] That is, the weight corresponding to the step filter s1(x,y) is β*α, the weight corresponding to the elliptic filter is β*(1-α), and the weight corresponding to the polynomial filter is 1-β. The weights adjusted in step 2 are these three weights.

[0068] Step 4: Perform the following steps on the image after image enhancement obtained in step 3: Figure 2 The indicated treatments gave 1 mg 2 , to achieve further image enhancement. Figure 2 The processing process is as follows:

[0069] 1) In the high-frequency extraction stage, the input image (i.e., the filtered image obtained in step 3) is first downsampled into RGB images of four scales to obtain four images of different sizes, such as Figure 2 The middle part of the image is located on the leftmost 4 layers of images of different sizes in each virtual box above the jump connection line, and then the high-frequency component extraction module (octave convolution is selected in this embodiment) is used to extract the high-frequency component of each layer of the image to obtain 4 high-frequency components of different scales. Of course, other high-frequency component extraction methods in the prior art can also be used, such as using hollow convolution to extract the high-frequency component of each layer of the image. This process is performed because the high-frequency component can enhance the edge details of the image.

[0070] 2) In the encoder stage, the input image (i.e., the filtered image obtained in step 3) is subjected to a 4-layer downsampling operation, i.e., Figure 2The processing of Stage 1, Stage 2, Stage 3 and Stage 4 on the left is to extract feature information (low-frequency components) of different scales respectively, and finally input to the decoder through a feature extraction module. The processing of Stage 1, Stage 2, Stage 3 and Stage 4 is similar, and each includes a feature extraction module and a feature expansion module.

[0071] 3) In the decoder stage, after Figure 2 The processing of the four-layer structure of Stage1, Stage2, Stage3 and Stage4 on the right. During the processing, Stage1 needs to jump-connect the output of the deepest layer of the encoder with the output of the encoder Stage4 to achieve feature expansion, and then fuse the feature expansion result with the high-frequency component of the same-scale layer, and then perform feature extraction (implemented by the feature extraction module of the decoder Stage1) for output; Stage2, Stage3 and Stage4 need to jump-connect the outputs of Stage1, Stage2 and Stage3 with the outputs of the corresponding stages of the encoder to achieve feature merging, and then fuse the feature merging result with the high-frequency component of the same-scale layer, and then perform feature extraction for output after fusion. Finally, the output of the decoder Stage4 is projected back to the original data space through a full connection to obtain Img. 2 In this process, the final Img 2 It is the result of the combined effect of high-frequency information and low-frequency information output by the encoder, which achieves effective expression of features and further enhances the image.

[0072] It should be noted that in this step, in the three stages of high frequency extraction, encoder and decoder, feature extraction, downsampling and other processing of four different scales are performed accordingly. The specific number of scales can be adjusted according to actual needs, such as three or five.

[0073] Step five: evaluate the image quality.

[0074] like Figure 5 As shown in the figure, the quality of the enhanced image is evaluated by calculating evaluation indicators such as structural similarity, peak signal-to-noise ratio, and minimum average error. Here, for the enhanced image obtained in step 4, α and β can be further adjusted based on the calculation results of structural similarity, minimum average error, and peak signal-to-noise ratio to ensure that the global contrast of the enhanced image is large, the structural similarity is high, and the local detail texture information is rich.

[0075] In summary, the present invention effectively solves the problem of enhancing degraded images collected in different scenes, especially solves the problem of enhancing degraded images under foggy, rainy, snowy and dusty weather conditions.

[0076] Specific implementation methods are given above, but the present invention is not limited to the described implementation methods. The basic idea of ​​the present invention lies in the above basic scheme. For ordinary technicians in this field, it does not take creative work to design various deformed models, formulas, and parameters according to the teachings of the present invention. Changes, modifications, substitutions, and variations of the implementation methods without departing from the principles and spirit of the present invention still fall within the scope of protection of the present invention.

Claims

1. An image enhancement method based on fusion parameter filter, characterized in that: The method includes: Acquire a scene image, extract three types of features, namely, grayscale features, texture features, and structural features, and determine the proportions of the three types of features; adjust the weights of the step filter, the elliptic filter, and the polynomial filter according to the proportions of the three types of features so that: when the difference between the proportions of any two types of features among the three types of features is less than a set difference threshold, the weights of the three filters are equal; otherwise, when the proportions of the grayscale features, texture features, and structural features are respectively the largest, the weights of the step filter, the elliptic filter, and the polynomial filter are correspondingly maximized; The scene image is input into a fusion parameter filter to obtain a filtered scene image; the fusion parameter filter is a filter obtained by weighted summing the three filters using the weights adjusted by the three filters.

2. The image enhancement method based on fusion parameter filter according to claim 1, characterized in that: Otherwise, when the proportions of grayscale features, texture features and structural features are respectively the smallest, the weights of the step filter, the elliptic filter and the polynomial filter are minimized accordingly; Or otherwise, the weights of the filters corresponding to the two types of features with smaller proportions are made equal; among them, the grayscale features, texture features and structural features correspond to the step filter, elliptical filter and polynomial filter respectively.

3. The image enhancement method based on fusion parameter filter according to claim 1, characterized in that: The method for determining the proportion of the three types of features is: normalize the vector representations corresponding to the three types of features respectively, convert the normalized vectors into numerical values ​​to obtain the numerical values ​​corresponding to the three types of features respectively; calculate the proportion of the numerical value corresponding to each type of feature to the sum of the corresponding numerical values ​​of the three types of features, and the proportion is the proportion corresponding to each type of feature.

4. The image enhancement method based on fusion parameter filter according to claim 3, characterized in that: The vector obtained by calculating the grayscale histogram of the scene image is the vector representation of the grayscale feature, the vector obtained by calculating the grayscale co-occurrence matrix of the scene image is the vector representation of the texture feature, and the vector obtained by calculating the gradient direction histogram of the scene image is the vector representation of the structural feature.

5. The image enhancement method based on fusion parameter filter according to claim 1, characterized in that: The method further includes: Downsampling the filtered scene images at different scales, extracting high-frequency components from the obtained downsampled images respectively, and obtaining high-frequency components at different scales; The encoder is used to extract feature information of different scales from the filtered scene image, and then the decoder is used to map the feature information of different scales output by the encoder back to the original data space. The output of each layer structure in the encoder is the result of skipping and fusing the corresponding scale feature information output by the encoder with the high-frequency component of the same scale, and then feature extraction is performed on the fused result.

6. The image enhancement method based on fusion parameter filter according to claim 2, characterized in that: The method of adjusting the weights of the three filters includes: if the proportion of grayscale features is greater than or equal to 0.5, the weight of the step filter is 0.5, and the weights of the elliptic filter and the polynomial filter are both 0.25; if the proportion of texture features is greater than or equal to 0.5, the weight of the elliptic filter is 0.5, and the weights of the step filter and the polynomial filter are both 0.25; if the proportion of structural features is greater than or equal to 0.5, the weight of the polynomial filter is 0.5, and the weights of the step filter and the elliptic filter are both 0.

25.

7. The image enhancement method based on fusion parameter filter according to claim 5, characterized in that: Octave convolution is used to extract high-frequency components.

8. The image enhancement method based on fusion parameter filter according to any one of claims 1 to 7, characterized in that: The acquired scene image is an image after denoising.

9. The image enhancement method based on fusion parameter filter according to any one of claims 1 to 7, characterized in that: The expression of the polynomial filter is s3(x,y)=Img*(x+y+γ) p Or s3(x,y)=(x+y+Img+γ) p , where s3 represents a polynomial filter, Img represents the intensity of the image pixel at the (x, y) position, γ represents an independent scalar, and p = 3.

10. The image enhancement method based on fusion parameter filter according to any one of claims 1 to 7, characterized in that: The method also includes: calculating the size of an evaluation index of the final image, the evaluation index including at least one of structural similarity, contrast sensitivity function, peak signal-to-noise ratio and minimum average error; evaluating the image enhancement effect according to the evaluation index, and further adjusting the weights corresponding to the three filters if the effect is not good.

Images