Infrared Image Enhancement Method Based on Circular Convolution and Multi-Clustering Space

Abstract

The present invention belongs to the field of image enhancement technology, and particularly relates to an infrared image enhancement method based on circular convolution and multi-clustering space, comprising the following steps: S1, establishing a multi-scale circular convolution module, establishing a multi-scale convolution module with a plurality of Gaussian convolution kernels having different radii, and connecting multiple multi-scale convolution modules in a densely connected manner to process image details; S2, using the K-means algorithm to perform clustering analysis on the image gray values, and selecting the initial clustering center by the maximum contribution rate method; S3, using a two-dimensional Gaussian function and a rotation matrix to construct the kernel function at the i th angle as the template kernel function to extract the contour information of the image at different angles. Compared with the existing image enhancement methods, the method of this embodiment outputs an image with clearer contours and richer detail information, and the quality of the original image has been significantly improved in terms of details, brightness, etc.

Description

Technical Field

[0001] The present invention belongs to the field of image enhancement technology, and particularly relates to an infrared image enhancement method based on circular convolution and multi-cluster space. Background Art

[0002] Improving the image quality, highlighting the detailed information of the image, and enhancing the image contrast by means of image enhancement are extremely important tasks for broadening the infrared imaging technology. The infrared image enhancement technology mainly includes: (1) The histogram enhancement method, which improves the image quality by changing the distribution of the infrared image histogram in the entire gray level range. However, due to factors such as merging the gray levels of the image and the single gray value mapping function, problems such as loss of image details, blurred contours, and blocking effects are prominent; (2) The image enhancement method based on image layering. This method layers the image according to the image features, and obtains a high-quality image by fusing the processed sub-layers enhanced by different methods, which can solve the problems of blurred image details and low contrast, but the image brightness is low; (3) The image enhancement method based on the transform domain. This method mainly converts the image gray value to the gradient domain or the wavelet domain for enhancement operations. This method can effectively improve the image details and texture details, but the overall quality of the image has not been well improved. Summary of the Invention

[0003] Aiming at the problems of low contrast, low brightness, and blurred contour details of the infrared image after processing by the existing image enhancement technology, the present invention provides an infrared image enhancement method based on circular convolution and multi-cluster space, which can enhance the image contrast while improving the image brightness, highlight the texture detail information of the image, and output high-quality images.

[0004] The technical solutions for achieving the above object include the following contents.

[0005] An infrared image enhancement method based on circular convolution and multi-cluster space, comprising the following steps:

[0006] S1, establishing a multi-scale circular convolution module, establishing a multi-scale convolution module with a plurality of Gaussian convolution kernels with different radii, and connecting multiple multi-scale convolution modules in a densely connected manner to process image details; the dense connection can make full use of the information of the image to achieve the effect of enhancing the image detail information;

[0007] S2, using the K-means algorithm to perform clustering analysis on the image gray values, and selecting the initial clustering center by the maximum contribution rate method; the traditional K-means randomly selects gray points as the clustering centers. However, the gray value p of the image P ij is usually relatively dense. If the randomly selected clustering centers coincide or are close to each other, the finally obtained clustering centers may not be the optimal solutions;

[0008] In S3, a two-dimensional Gaussian function and a rotation matrix γ are adopted. i Construct the kernel function at the i-th angle as the template kernel function to extract the contour information of the image at different angles. The image contours and detail images extracted by the multi-angle filter are fused with the image processed by multi-cluster spatial equalization according to different weights to obtain a high-quality image with rich details, clear contours, and high contrast.

[0009] Furthermore, the convolutional layer described in S1 is n layers. The previous layer uses Adaptive Histogram Equalization (AHE) as the input for all subsequent layers to solve for a higher-quality image.

[0010] Furthermore, the convolutional layer described in S1 is 2 layers. The original image I is fused with the feature F1 extracted by the first convolutional layer and used as the input I1 for the second convolutional layer. The second convolutional layer uses Adaptive Histogram Equalization (AHE) to perform contrast stretching on the feature F1 extracted by the first convolution to obtain F2, and F2 is fused with the original image I and F1 to solve for a higher-quality image I2.

[0011] Furthermore, the initial cluster center calculation in S2 is as follows. Assume that the gray point p in the image P i and other gray levels p j The cosine distance D pi is

[0012]

[0013] The gray value p in the image P i and other gray levels p in the image j The minimum value of the distance is defined as d i , that is

[0014] d i = min(D pi ) (2), The gray value p in the image P i and the cumulative number of the distance to other gray levels being d i is defined ρi as, that is

[0015] ρ i = size(find(D pi == d i )) (3), Then, the contribution rate i of the gray value p in the image P is

[0016]

[0017] Arrange them in descending order, that is,

[0018]

[0019] Select the top K gray values with the largest contribution rate, that is The gray value of the corresponding image P As the initial clustering center.

[0020] Furthermore, in S3, the rotation matrix γ at the i-th angle i is

[0021]

[0022] If P(x, y) is a point on the kernel function in the X-Y coordinate system, after rotation processing, the point in the U-V coordinate system at the i-th angle is:

[0023]

[0024] In order to uniformly filter the image background and extract the detailed information of the image, that is, the response of the background area (non-contour and detailed information) after template filtering is 0, define the filter kernel function at the i-th angle as

[0025]

[0026] Where belongs to the neighborhood N is the neighborhood The number of gray values in represents rounding, Lap is the Laplacian operator, which is used to extract the detailed information of each angle of the image. Description of the Drawings

[0027] Figure 1 It is a flowchart of the infrared image enhancement method based on circular convolution and multi-clustering space provided by the invention;

[0028] Figure 2 It is a schematic diagram of the multi-scale circular convolution module of this embodiment;

[0029] Figure 3 It is a schematic diagram of the multi-clustering space hybrid equalization module of this embodiment;

[0030] Figure 4 It is a comparison diagram of the original image and the multi-scale circular convolution enhancement of this embodiment;

[0031] Figure 5The original image of this embodiment and the comparison image after multi-scale cyclic convolution and multi-cluster spatial equalization processing;

[0032] Figure 6 The original image of this embodiment and the comparison image after processing with the Laplace operator and the multi-angle combined filter;

[0033] Figure 7 The comparison image between the original image of this embodiment and the enhanced image using the processing of this embodiment.

[0034] Figure 8 The comparison image processed by using multiple different existing technologies and the method of this embodiment for the same infrared grayscale image. Detailed implementation manners

[0035] The present invention will be specifically described below in conjunction with embodiments.

[0036] See Figures 1 to 8 In this embodiment, for enhancing a single infrared grayscale image, a multi-scale cyclic convolution module is designed. See Figure 2 This module includes two convolutional layers. Each convolutional layer contains 6 Gaussian convolutional kernels of size 3*3 with convolution radii of 0.1, 0.3, 0.5, 0.7, 0.9, and 1 respectively. The original image I is fused with the feature F1 extracted by the first convolutional layer and used as the input I1 of the second convolutional layer. The second convolutional layer uses Adaptive Histogram Equalization (AHE) to perform contrast stretching on the feature F1 extracted by the first convolution to obtain F2. F2 is fused with the original image I and F1 to solve for a higher-quality image I2. The two convolutional layers of this embodiment are connected. Figure 4 It shows that the output image I2 on the right effectively enhances the detail information of the image and improves the visual effect of the image compared with the original image on the left;

[0037] Based on the above image I2, an improved K-means algorithm is used to perform clustering analysis on the image gray values, and then different equalization methods are used for enhancement according to their characteristics in different clustering spaces. The traditional K-means randomly selects gray points as the clustering centers. However, the gray value p of the image P ij is usually densely distributed. If the randomly selected clustering centers coincide or are close to each other, the finally solved clustering centers will not be the optimal solutions. Therefore, when performing image clustering analysis, the maximum contribution rate method is first used to select the initial clustering centers. The specific steps are as follows.

[0038] 1) Initial clustering center calculation

[0039] Assume that the gray point p in the image P iThe cosine distance D j between it and other gray levels p pi is

[0040]

[0041] The gray level value p in image P i and other gray levels p in the image j The minimum value of the distance is defined as d i That is

[0042]

[0043] For the image P the cumulative number of gray level values pi with a distance of di from other gray level values is defined ρi as, that is

[0044]

[0045] Then, the contribution rate i of the gray level value p in image P is

[0046]

[0047] Sorted in descending order, that is,

[0048]

[0049] Select the top K gray level values with the largest contribution rate, that is The corresponding gray level values of image P are used as the initial clustering centers

[0050] 2) Gray level category division

[0051] The gray level values closest to the clustering center are grouped into one category, that is,

[0052]

[0053] Among them, represents the category i at the m-th iteration, s p represents any gray level value of image P, represents the clustering center with the largest contribution rate, represents the clustering center with non-maximum largest contribution rate,

[0054] 3) Iterative accuracy evaluation

[0055] Calculate the iterative accuracy of each sample in different categories, that is

[0056]

[0057] Among them, s p represents any gray value in the image, represents the clustering center. λ ph is the weight. If s p belongs to the category with as the clustering center, its value is 1; otherwise, the value is 0, that is

[0058]

[0059] If the iteration accuracy does not meet the set threshold, then update the clustering center,

[0060] 4) Update of clustering center

[0061] Solve the mean values of different categories and use them as the new clustering centers. Repeat steps two and three until the iteration accuracy meets the set threshold. The clustering center of the h-th category in the (m + 1)-th iteration is,

[0062]

[0063] Among them, T represents the number of gray levels in the category with as the clustering center,

[0064] 5) Equalization

[0065] After performing clustering analysis on the gray values of the image, divide the gray values of the image into multiple clustering spaces, and determine the histogram range of the image after equalization according to the proportion of the image information contained in each clustering space, so as to achieve the purpose of enhancing the image. Assume that the number of gray levels of image P is N, and the number of gray levels contained in a certain clustering space q is M q , then the histogram range of this clustering space after equalization is:

[0066]

[0067] Among them, represents rounding, represents the maximum value of the histogram of clustering space q, represents the minimum value of the histogram of clustering space q, which is determined by the maximum value of the histogram of the previous clustering space q - 1, that is

[0068]

[0069] This step divides the image into three clustering spaces. Since the first clustering space mainly contains the detailed information of the darker regions in the image, according to the histogram range of the clustering space q, the traditional histogram equalization method is used to perform equalization operations within the specified range; since the second clustering space includes most of the information in the image, only the gamma correction method is used to perform brightness correction on the image; since the third clustering space includes a small amount of contour information in the image, the linear correction method is used to perform linear stretching within the specified interval range. Figure 5 The central part is Figure 4 The output image I2 in

[0070] Based on the image I3 processed by the above multi-clustering space equalization, the two-dimensional Gaussian function and the rotation matrix γ are used i to construct the kernel function at the i-th angle as the template kernel function to extract the contour information of the image at different angles. The specific steps are as follows: the rotation matrix γ at the i-th angle i is

[0071]

[0072] If P(x, y) is a point on the kernel function in the X-Y coordinate system, the point in the U-V coordinate system at the i-th angle after rotation processing is:

[0073]

[0074] In order to uniformly filter the image background and extract the detailed information of the image, that is, the response of the background area (non-contour and detailed information) after template filtering is 0, the filter kernel function at the i-th angle is defined as

[0075]

[0076] where belongs to the neighborhood N is the number of gray values in the neighborhood represents taking the integer, Lap is the Laplace operator, which is used to extract the detailed information of each angle of the image, Figure 6 The middle part is the contour map extracted by the existing Laplace operator. The left side is the original image, and the right side is the contour and detailed information of the image extracted by the multi-angle combination filter used in this embodiment. It can be seen that the image contour extracted by the method of this embodiment is clearer and the detailed information is more abundant;

[0077]

[0077] Finally, a linear fusion strategy is adopted to fuse two images of the same object with different features together to improve the quality of infrared images. For the processing results, see Figure 7 , where the original image is on the left and the final image processed in this embodiment is on the right. Figure 8 The final image processed in this embodiment is on the far right, the original image is on the far left, and the middle images are the results processed by the existing four image enhancement methods. It can be seen that, compared with the existing image enhancement methods, the image output by the method of this embodiment has a clearer contour and richer detail information, and there is a significant improvement in details, brightness, etc. compared with the original image.

Claims

1. An infrared image enhancement method based on circular convolution and multi-cluster space, characterized in that The method includes the following steps: S1. Establish a multi-scale recurrent convolutional module, which includes convolutional layers. A multi-scale convolutional module is established with Gaussian convolutional kernels of several different radii, and multiple multi-scale convolutional modules are connected in a densely connected manner to process image details; S2. Use the K-means algorithm to perform clustering analysis on the gray values of the image, and select the initial clustering center by the maximum contribution rate method; S3. Use a two-dimensional Gaussian function and a rotation matrix γ i Construct the kernel function at the i-th angle as the template kernel function to extract the contour information of the image at different angles; In S3, the rotation matrix γ for the i-th angle i is If P(x, y) is a point on the kernel function in the X-Y coordinate system, the point in the U-V coordinate system at the i-th angle after rotation processing is as follows: To uniformly filter the image background while extracting the detailed information of the image, that is, the response of the background area (non-contour and detailed information) after template filtering is 0, the filter kernel function at the i-th angle is defined be Among them, belongs to the neighborhood N is the neighborhood the number of gray values in it, denotes rounding, Lap is the Laplacian operator, which is used to extract the detail information of the image at various angles.

2. The infrared image enhancement method based on circular convolution and multi-clustering space according to claim 1, wherein In S1, the convolutional layer is n layers, and the previous layer uses adaptive histogram equalization as the input of all subsequent layers to solve a higher-quality image.

3. The infrared image enhancement method based on circular convolution and multi-clustering space according to claim 1, wherein In S1, the convolutional layer is 2 layers. The original image I is fused with the feature F1 extracted by the first convolutional layer and used as the input I1 of the second convolutional layer. The second convolutional layer uses adaptive histogram equalization to perform contrast stretching on the feature F1 extracted by the first convolutional layer to obtain F2. F2 is fused with the original image I and F1 to solve a higher-quality image I2.

4. The infrared image enhancement method based on circular convolution and multi-clustering space according to claim 1, wherein The initial cluster centers described in S2 are calculated as follows. Assume that the gray point p in the image P i and other grays p j the cosine distance between is Gray value p in image P i and other gray value p in the image j The minimum value of the distance is defined as d i That is Image P The gray value p i has a distance d from other gray values i and the cumulative number is defined as ρ i That is Then, the contribution rate of the gray value p in the image P i is as follows Arrange them in descending order, that is, Select the top K gray values with the largest contribution rate, that is The gray values of the corresponding image P As the initial clustering centers.

Images