Underwater Image Mosaic Method with Feature Point Matching Integrating Luminance Region Information

Abstract

The present invention discloses an underwater image stitching method for feature point matching integrating brightness region information, including: 1) extraction of the brightness region of the underwater image; 2) feature point matching integrating brightness region information; 2.1) rough matching of the underwater image based on the brightness significant region; 2.2) fine matching of the underwater image based on the scale-invariant feature transform algorithm; 3) completing image stitching using the weighted fusion method. Aiming at the problem of poor quality of underwater images, the present invention adds rough matching based on the brightness significant region before the feature point matching of the traditional scale-invariant feature transform (SIFT) algorithm, which conforms to the characteristic that the overall underwater image is dark and the brightness region is easy to distinguish, makes up for the defects of difficult and unstable feature detection of underwater images, and improves the quality of underwater image stitching.

Description

Technical Field

[0001] The present invention relates to an underwater image stitching method, and particularly to an underwater image stitching method for feature point matching integrating luminance region information. Background Art

[0002] Underwater images refer to images located below the water surface captured by underwater image devices such as unmanned submersibles or underwater vehicles. In order to obtain a broader perspective or more information, image stitching technology is widely used. Since light undergoes severe attenuation, scattering, and absorption when propagating in water, the photos taken underwater are often of lower quality than those taken in air, and the stitched image effect is also not good. Therefore, a method that can be widely applied to underwater image stitching is urgently needed.

[0003] Image stitching is to register two or more images that overlap with each other, and then stitch them into a picture with a broader perspective. There are generally the following categories of image stitching methods: region-based registration methods and feature-based registration methods.

[0004] The region-based registration method starts from the gray values of the images to be stitched. Hu Shejiao et al. from Hefei University of Technology proposed an improved algorithm for image stitching based on gray correlation (Hu Shejiao, Tu Guilin, Jiang Ping. An improved algorithm for image stitching based on gray correlation [J]. Journal of Hefei University of Technology (Natural Science), 2008(06):863 - 865.). By calculating the gray average value on the selected feature block and the difference between each pixel and the average value, and then selecting a certain threshold to reduce the image search range, and finally using the gray correlation method to match the two images to be stitched, but its gray matching is not accurate enough, and seams are inevitably formed in the overlapping part of the corresponding regions. Regarding the accuracy of gray matching, Tian Weifeng et al. from the Chinese Flight Test and Research Institute proposed an image stitching algorithm based on gray correlation and regional features (Tian Weifeng, Chen Bei, Liu Qian. An image stitching algorithm based on gray correlation and regional features [J]. Electronic Design Engineering, 2011, 19(03):184 - 186.). This algorithm uses the method of gray histogram equalization to reduce the gray difference caused by different lighting conditions, calculates the gray average value and the absolute difference between each pixel and the average value on the selected feature block, and finally introduces a smoothing factor to achieve seamless stitching. However, the application range of this method is relatively narrow and cannot achieve a good stitching effect. In summary, the region-based registration method does not detect the features in the image, but directly matches according to the gray values or shapes of the regions, has a relatively narrow application range, is prone to generating seams, and has a poor stitching effect.

[0005] The most representative feature-based method is the Scale-invariant Feature Transform (SIFT) algorithm. In 2021, Gu Zishan et al. from Nanjing Institute of Technology proposed a panoramic image stitching method based on binocular cameras (Publication No.: CN 113191954 A). This method uses SIFT feature point detection, then uses an improved Random Sample Consensus (RANSAC) algorithm for feature matching, and then automatically corrects the image brightness difference to obtain a relatively clear and natural panoramic image. However, this method is only applicable to image stitching in the air and cannot achieve good stitching for underwater images with low quality. In view of the characteristics of poor underwater image quality, in 2021, Zhang Senlin et al. from Zhejiang University proposed an underwater image stitching method based on multi-scale image fusion and SIFT features (Grant Publication No.: CN 111260543 B). This method registers underwater preprocessed images through an improved SIFT algorithm, significantly improving the quality of underwater images and enhancing the subsequent image registration effect. However, the quality of underwater photos is low, feature point detection is unstable, and it cannot guarantee the improvement of feature point matching effect in the later stage under ordinary preprocessing. To sum up, the feature-based image stitching method is applicable to areas with significant local features, but it is difficult and unstable to detect features for underwater images with poor quality, and it is prone to false stitching phenomena.

[0006] To sum up, the limitations of the region-based image stitching method are relatively obvious, seams are easily generated, and the stitching effect is poor. The pure region-based image stitching method is rarely used at present; the feature-based image stitching method is widely used at present. Compared with the region-based matching method, its advantage lies in being generally applicable to situations where local structural information is more significant than gray information and can handle complex deformations between images. However, feature detection is difficult and unstable. For underwater images with low quality, feature extraction is relatively difficult and has certain limitations. Therefore, a stitching method applicable to underwater images needs to be proposed urgently. Summary of the Invention

[0007] The object of the present invention is to overcome the defects of the prior art and provide an underwater image stitching method that integrates brightness region information for feature point matching, which can effectively improve the quality of underwater image stitching.

[0008] The object of the present invention is achieved as follows: An underwater image stitching method that integrates brightness region information for feature point matching, characterized by including the following steps:

[0009] Step 1) Extraction of the brightness region of the underwater image;

[0010] Step 2) Feature point matching integrating luminance region information;

[0011] Step 2.1) Coarse matching of underwater images based on luminance significant regions;

[0012] Step 2.2) Fine matching of underwater images based on the Scale-Invariant Feature Transform (SIFT) algorithm;

[0013] Step 3) Complete image stitching using the weighted fusion method.

[0014] Further, the specific steps of Step 1) include: extracting the highly significant regions of two pictures based on features; obtaining underwater images with the assistance of artificial light sources, and the luminance significant regions have two features: high luminance and connectivity;

[0015] Regarding the high-luminance feature, first obtain the grayscale image of the underwater image, set the luminance threshold. When the grayscale value is higher than this luminance threshold, take 1, and when it is lower than this luminance threshold, take 0. That is, the pixel grayscale values in the obtained luminance significant suspected region S h satisfy Equation (1),

[0016]

[0017] where Y(x, y) is the grayscale value of the underwater image at coordinates (x, y), m is the maximum pixel grayscale value in the image, L is the ambient light intensity, and all pixel points in the image that satisfy Equation (1) constitute the luminance significant suspected region S h ;

[0018] Regarding the connectivity feature, according to the pixel grayscale values, group the pixels with similar grayscale values in the image into the same category, and solve the optimization objective function shown in Equation (2):

[0019]

[0020] In the formula, the number of categories N = 8, I i is the pixel in the image, C k is the clustering center, and the initial value is selected at equal intervals in the image to obtain the classification result. Select the region corresponding to the brightest category, and this region is S c , set all pixels in this region to 1, and the remaining regions to 0;

[0021] Considering both the high luminance and connectivity features of the luminance region, obtain the pixel set S of the luminance region, as shown in Equation (3):

[0022] S = {I(x, y)|S h (x, y) = 1 & S c (x, y) = 1} (3)

[0023] Among them, S represents the luminance region, I(x, y) represents the pixel value at the coordinate (x, y) in the luminance region, and S h (x, y) represents the pixel value at the coordinate (x, y) in the region S h ; S c (x, y) represents the pixel value at the coordinate (x, y) in the region S c ; By this method, the luminance significant regions S of the two images to be stitched are extracted respectively A , S B ; If the luminance significant region is not extracted from at least one picture, first perform image preprocessing on the underwater image, and then directly execute step 2.2).

[0024] Furthermore, the specific steps of step 2.1) are as follows: The luminance significant regions S of the two images are obtained from step 1 A , S B , and the rough matching process based on the luminance significant region of the image is as follows:

[0025] In the two images to be stitched, use S A_hb to represent a certain luminance significant region of image A, and use S B_hb to represent a certain luminance significant region in image B. Through formula (4), the registration correlation μ of the selected luminance significant regions of the two images can be obtained

[0026]

[0027] Among them, and respectively represent the average pixel gray values of pictures A and B, and S A_hb (x, y) represents the pixel gray value corresponding to the coordinate (x, y) in the luminance significant region S of picture A A_hb ; S B_hb (x, y) represents the pixel gray value corresponding to the coordinate (x, y) in the luminance significant region S of picture B B_hb ; exp(-(N b / N s )) is the weight, N b is the number of pixels in the larger area of the two luminance regions to be registered, and N s is the number of pixels in the smaller area of the two luminance regions to be registered;

[0028] Assume that m luminance significant regions are extracted from picture A, which are respectively represented as S A1 ,..., S Am ; n luminance significant regions are extracted from picture B, which are respectively represented as S B1 ,..., S Bn;Use Equation (4) to solve the registration correlation μ for the m luminance significant regions of Image A and the n luminance significant regions of Image B respectively, and select the two luminance significant regions corresponding to the maximum value of the registration correlation μ, which are the luminance significant regions corresponding to the registration of Image A and Image B. Exclude the remaining regions to complete the rough matching of underwater images based on luminance significant regions.

[0029] Further, the specific steps of step 2.2) include:

[0030] 1) Feature detection in the scale space: Search for stable features at all possible scales in the scale space from the underwater image, and perform a convolution operation on the input image,

[0031] L(x, y, σ) = G(x, y, σ) * I(x, y) (5)

[0032] where G(x, y, σ) is the Gaussian convolution kernel, and I(x, y) represents the original image; the scale space D(x, y, σ) of the image is obtained by combining the Gaussian difference operation on the basis of the convolution image L(x, y, σ), as shown in the following formula,

[0033] D(x, y, σ) = (G(x, y, kσ) - G(x, y, σ)) * I(x, y) = L(x, y, kσ) - L(x, y, σ) (6)

[0034] Detect the local maxima and minima in the neighborhood to identify the potential extreme points of the underwater image;

[0035] 2) Key point localization: Further refine the candidate feature points in the scale space extracted from the underwater image to adapt to the position, scale, and principal curvature ratio; the Taylor expansion of the Gaussian difference function is:

[0036]

[0037] Take the derivative of X and set the equation equal to 0 to determine the position of the candidate feature points in the underwater image, and then use the Gaussian difference function at the candidate feature points to remove the unstable features with low contrast in the underwater image;

[0038] Obtain the Hessian matrix H at the key points,

[0039]

[0040] The eigenvalues of the matrix are α and β, and the sum and product of the above eigenvalues can be calculated by the trace and determinant of H respectively:

[0041]

[0042]

[0043] where (r + 1) 2 / r increases as r increases and reaches a minimum if the eigenvalues α and β are equal; the ratio of the principal curvatures is checked against a certain threshold r through the following equation (11), and candidate feature points that are poorly located along the underwater image edges are discarded.

[0044]

[0045] If the formula is satisfied, the key points are retained; otherwise, they are eliminated.

[0046] 3) Orientation assignment: Extract the scale of the feature points from the underwater image, select the Gaussian-smoothed image L with the closest scale. For each image sample L(x, y) at this scale, use the pixel differences to calculate the gradient magnitude m(x, y) and direction θ(x, y).

[0047]

[0048]

[0049] Form an orientation histogram from the gradient directions within the region around the feature points in the underwater image. The peak in the orientation histogram corresponds to the dominant direction of the local gradient of the underwater image feature points.

[0050] 4) Key point description: First, calculate the gradient magnitude and direction of the underwater image and weight them using a Gaussian function within the region around the feature point positions in the underwater image; the above sampling is accumulated into an orientation histogram, which summarizes the content on the image sub-region, and the length of the arrow corresponds to the sum of the gradient magnitudes near that direction; after that, the feature vector of the underwater image is further normalized and described as a unit length.

[0051] 5) Feature matching: Remove the mismatched feature points in the SIFT algorithm through the RANSAC algorithm.

[0052] The present invention adopts the above technical solutions. Compared with the prior art, the beneficial effects are as follows: The present invention extracts the significantly bright regions of two underwater images; secondly, performs rough matching of the underwater images based on the significantly bright regions of the images; then, uses the scale-invariant feature transform algorithm to complete the fine matching of the underwater images through feature point matching; finally, uses the weighted fusion method to complete the image stitching; a rough matching based on the significantly bright regions is added before the feature point matching of the traditional SIFT algorithm, which conforms to the characteristic that the overall underwater image is relatively dark and the bright regions are easy to distinguish, makes up for the defects of difficult and unstable feature detection of underwater images, and improves the quality of underwater image stitching. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is a flowchart of the present invention. Specific implementation mode

[0054] As Figure 1 shown in the underwater image stitching method of feature point matching integrating luminance region information, including the following steps:

[0055] Step 1) Extraction of underwater image luminance region;

[0056] Extract the highly significant regions of two pictures based on two features; The diver carries a lighting device and takes two photos with a small displacement difference underwater using a camera as the pictures to be fused. Intuitively, the luminance significant region has two features: high luminance and connectivity.

[0057] Regarding the high-luminance feature, first obtain the grayscale image of the underwater image, set the luminance threshold. When the grayscale value is higher than the luminance threshold, take 1, and when it is lower than the luminance threshold, take 0, that is, the pixel grayscale values in the obtained luminance significant suspected region S h satisfy Equation (1),

[0058]

[0059] where Y(x, y) is the grayscale value of the underwater image at coordinates (x, y), m is the maximum pixel grayscale value in the image, L is the ambient light intensity, and all pixel points in the image that satisfy Equation (1) constitute the luminance significant suspected region S h .

[0060] Regarding the connectivity feature, according to the pixel grayscale values, divide the pixels with similar grayscale values in the image into the same category, and solve the optimization objective function shown in Equation (2):

[0061]

[0062] In the formula, the number of categories N = 8, I i is the pixel in the image, C k is the clustering center, and the initial value is selected at equal intervals in the image. Obtain the classification result, select the region corresponding to the brightest category, and this region is S c , set all pixels in this region to 1, and the rest of the region to 0.

[0063] Comprehensively considering the two features of high luminance and connectivity of the luminance region, obtain the pixel set S of the luminance region, as shown in Equation (3):

[0064] S = {I(x, y)|S h (x, y) = 1 & S c (x, y) = 1} (3)

[0065] where S represents the luminance region, I(x, y) represents the pixel value at coordinates (x, y) in the luminance region, Sh (x, y) represents the pixel value at the coordinate (x, y) in the region S h in the image, and S c (x, y) represents the pixel value at the coordinate (x, y) in the region S c in the image. By this method, the luminance significant regions S of the two images to be stitched are extracted respectively A , S B . If at least one image fails to extract the luminance significant region, first perform image preprocessing on the underwater image, and then directly execute step 2.2 to perform feature point matching using the traditional SIFT algorithm

[0066] Step 2) Feature point matching for integrating luminance region information

[0067] Step 2.1) Coarse matching of underwater images based on luminance significant regions

[0068] From step 1), the luminance significant regions S of the two images have been obtained A , S B . The process of coarse matching based on the luminance significant regions of the images is as follows

[0069] In the two images to be stitched, use S A_hb to represent a certain luminance significant region of image A, and use S B_hb to represent a certain luminance significant region of image B. Through formula (4), the registration correlation μ of the selected luminance significant regions of the two images can be obtained

[0070]

[0071] where and represent the average pixel grayscale values of pictures A and B respectively, and S A_hb (x, y) represents the pixel grayscale value corresponding to the coordinate (x, y) in the luminance significant region S of picture A A_hb in the image, and S B_hb (x, y) represents the pixel grayscale value corresponding to the coordinate (x, y) in the luminance significant region S of picture B B_hb in the image. exp(-(N b / N s )) is the weight, N b is the number of pixels in the larger area of the two luminance regions to be registered, and N s is the number of pixels in the smaller area of the two luminance regions to be registered, representing the area relationship between the two luminance regions to be sought. The greater the difference in area, the smaller the weight, that is, the lower the possibility of matching between the two regions

[0072] From formula (4), it can be obtained that if the value of μ is larger, then the luminance significant region S A_hb and SB_hb The higher the registration accuracy. In the actual operation process, usually the number of highlighted significant regions extracted from two underwater images is not the same. Suppose the number of brightness significant regions extracted from picture A is m, which are respectively represented as S A1 ,..., S Am ; the number of brightness significant regions extracted from picture B is n, which are respectively represented as S B1 ,..., S Bn . The registration correlation μ is respectively solved for the m brightness significant regions of picture A and the n brightness significant regions of picture B by using formula (4). The two brightness significant regions corresponding to the maximum value of the registration correlation μ are selected, which are the brightness significant regions corresponding to the registration of image A and image B. The remaining regions are removed to complete the rough matching of underwater images based on the brightness significant regions.

[0073] Step 2.2) Fine matching of underwater images based on the scale-invariant feature transform algorithm;

[0074] 1) Feature detection in the scale space: Search for stable features at all possible scales in the scale space from the underwater image, and perform a convolution operation on the input image,

[0075] L(x, y, σ) = G(x, y, σ) * I(x, y) (5)

[0076] where G(x, y, σ) is the Gaussian convolution kernel, and I(x, y) represents the original image. The scale space D(x, y, σ) of the image is obtained by combining the Gaussian difference operation on the basis of the convolution image L(x, y, σ), as shown in the following formula,

[0077] D(x, y, σ) = (G(x, y, kσ) - G(x, y, σ)) * I(x, y) = L(x, y, kσ) - L(x, y, σ) (6)

[0078] Detect the local maxima and minima in the neighborhood to identify the potential extreme points of the underwater image, and these extreme points do not change with scale and direction.

[0079] 2) Key point localization: The candidate points solved in the above process are not all feature points, and they need to be screened. The candidate feature points in the scale space extracted from the underwater image are further refined to adapt to the position, scale, and principal curvature ratio. The Taylor expansion of the Gaussian difference function is:

[0080]

[0081] Take the derivative of X and set the equation equal to 0 to determine the position of the candidate feature points of the underwater image, and then use the Gaussian difference function at the candidate feature points to remove the unstable features with low contrast in the underwater image.

[0082] In addition, the above Difference of Gaussians function itself has a strong boundary effect in underwater images. This boundary effect is unstable to noise. Obtain the Hessian matrix H at the key points,

[0083]

[0084] The eigenvalues of the matrix are α and β. The sum and product of the above eigenvalues can be calculated from the trace and determinant of H respectively:

[0085]

[0086]

[0087] where (r + 1) 2 / r increases with r and reaches the minimum value if the eigenvalues α and β are equal. Therefore, we can apply the following equation to check whether the ratio of the principal curvatures is lower than a certain threshold r and discard the candidate key points that are poorly located along the edges of the underwater image,

[0088]

[0089] If the formula is satisfied, the key points are retained; otherwise, they are eliminated.

[0090] 3) Direction assignment: When consistent directions are assigned to each key point according to the local characteristics of the underwater image, rotational invariance can be further achieved. Extract the scale of the key points from the underwater image, and select the Gaussian-smoothed image L with the closest scale to achieve scale invariance; for each image sample L(x, y) at this scale, use the pixel difference to calculate the gradient magnitude m(x, y) and direction θ(x, y),

[0091]

[0092]

[0093] A direction histogram can be formed from the gradient directions within the region around the key points of the underwater image. The peak in the direction histogram corresponds to the dominant direction of the local gradient of the underwater image key points.

[0094] 4) Key point description: First, calculate the gradient magnitude and direction of the underwater image, and weight them with a Gaussian function within the region around the key point positions in the underwater image. The above sampling is accumulated into a direction histogram, which summarizes the content on the image sub-region. The length of the arrow corresponds to the sum of the gradient magnitudes near that direction; after that, the feature vector of the underwater image is further normalized and described as a unit length.

[0095] 5) Feature matching: Remove the mismatched feature points in the SIFT algorithm through the RANSAC algorithm to improve problems such as uneven illumination, low contrast, and obvious noise in underwater images, and solve the problems of low matching efficiency and poor robustness of underwater images.

[0096] Step 3) Complete image stitching using the weighted fusion method;

[0097] Directly stitching two underwater pictures according to the corresponding feature points and regions will produce obvious seams. In order to make the stitched picture smoother, a weighted fusion algorithm is used to achieve image fusion. Let I 1 (x 1 , y 1 ) be the pixel value of image A at point (x 1 , y 1 ), I 2 (x 2 , y 2 ) be the pixel value of image B at point (x 2 , y 2 ), and I 3 (x 3 , y 3 ) be the pixel value of the fused image at (x 3 , y 3 ). w 1 and w 2 are the weight values of the two underwater images in the overlapping area respectively, and the fusion formula is as follows:

[0098]

[0099] Through this fusion formula, the finally stitched underwater image is obtained.

[0100] In the present invention, the significantly bright regions of two underwater images are extracted, and then rough matching of underwater images is performed based on the significantly bright regions of the images; fine matching of underwater images is completed using feature point matching based on the scale-invariant feature transform algorithm; finally, image stitching is completed using the weighted fusion method. Aiming at the problem of poor quality of underwater images, rough matching based on the significantly bright region is added before the feature point matching of the traditional scale-invariant feature transform (SIFT) algorithm, which conforms to the characteristic that the overall underwater image is dark and it is easy to distinguish the bright region, makes up for the defects of difficult and unstable feature detection of underwater images, and improves the quality of underwater image stitching.

[0101] The present invention is not limited to the above embodiments. Based on the technical solutions disclosed in the present invention, those skilled in the art can make some substitutions and deformations of some technical features without creative labor according to the disclosed technical content, and these substitutions and deformations are all within the protection scope of the present invention.

Claims

1. An underwater image stitching method based on feature point matching integrating luminance region information, characterized in that, it includes the following steps: Step 1) Extraction of the underwater image luminance region; Step 2) Feature point matching integrating luminance region information; Step 2.1) Coarse matching of the underwater image based on the luminance significant region; Step 2.2) Fine matching of the underwater image based on the scale-invariant feature transform algorithm; Step 3) Completing image stitching using the weighted fusion method; Step 1) specifically includes: extracting the highly significant regions of two pictures based on two features; obtaining the underwater image with the assistance of an artificial light source, and the luminance significant region has two features: high luminance and connectivity; For high-brightness features, first obtain the grayscale image of the underwater image, set the brightness threshold. If the grayscale value is higher than the brightness threshold, take 1; if it is lower than the brightness threshold, take 0. That is, the pixel grayscale values in the obtained significantly suspected brightness region S h satisfy Equation (1). Where Y(x, y) is the gray value of the underwater image at the coordinate (x, y), m is the maximum gray value of the pixels in the image, L is the ambient light intensity, and all the pixel points in the image that satisfy Equation (1) constitute the significantly bright suspected area S h ; For the connectivity feature, according to the pixel gray value, the pixels with similar gray values in the image are classified into the same category, and the optimization objective function shown in Equation (2) is solved: where the number of categories N = 8, I i is the pixel in the image, C k is the clustering center, and the initial value is selected at equal intervals in the image to obtain the classification result. The area corresponding to the brightest category is selected, and this area is S c . All pixels in this area are set to 1, and the remaining areas are set to 0; Taking into account the two features of high luminance and connectivity of the luminance region, the pixel set S of the luminance region is obtained, as shown in Equation (3): S = {I(x, y)|S h (x, y) = 1 & S c (x, y) = 1} (3) Among them, S represents the luminance region, I(x, y) represents the pixel value at the coordinate (x, y) in the luminance region, and S h (x, y) represents the pixel value at the coordinate (x, y) in the region S h ; S c (x, y) represents the pixel value at the coordinate (x, y) in the region S c ; The luminance significant regions S of the two images to be stitched are extracted respectively by this method A , S B ; If the luminance significant region is not extracted in at least one picture, first perform image preprocessing on the underwater image, and then directly execute Step 2.2).

2. The underwater image stitching method based on feature point matching integrating luminance region information according to Claim 1, characterized in that, The specific steps of step 2.1) include: two significantly bright regions S of the images are obtained from step 1) A , S B . The rough matching process based on the significantly bright regions of the images is as follows: In two images to be stitched, use S A_hb to represent a certain significantly bright region of Image A, and use S B_hb to represent a certain significantly bright region of Image B. Through formula (4), the registration correlation μ of the selected significantly bright regions of the two images can be obtained. Among them, and respectively represent the average pixel gray value of Picture A and Picture B, S A_hb (x, y) represents the pixel gray value corresponding to the coordinate (x, y) in the significant brightness area S A_hb of Picture A, and S B_hb (x, y) represents the pixel gray value corresponding to the coordinate (x, y) in the significant brightness area S B_hb of Picture B; exp(-(N b / N s )) is the weight value, N b is the number of pixels of the larger area among the two brightness areas to be registered, and N s is the number of pixels of the smaller area among the two brightness areas to be registered; Suppose that m luminance significant regions are extracted from Image A, which are respectively denoted as S A1 ,..., S Am ; n luminance significant regions are extracted from Image B, which are respectively denoted as S B1 ,..., S Bn ; The registration correlation μ is respectively solved for the m luminance significant regions of Image A and the n luminance significant regions of Image B using Equation (4). The two luminance significant regions corresponding to the maximum value of the registration correlation μ are selected, which are the luminance significant regions corresponding to the registration of Image A and Image B. The remaining regions are removed to complete the rough matching of underwater images based on luminance significant regions.

Images