Self-adaptive underwater image enhancement method based on pyramid fusion

Abstract

The invention provides a self-adaptive underwater image enhancement method based on pyramid fusion. The method mainly comprises the following steps: compensating and balancing a blue-green channel for inputting an underwater image, and iteratively compensating a red channel by using the balanced blue-green channel; weakening noise of the underwater image after color correction by using a guide filter, and then stretching pixel values of red, green and blue channels and a brightness channel in a global range; extracting and enhancing details of the stretched image through multi-scale filtering to obtain a first result, and performing histogram equalization of dynamic contrast limitation on the image to obtain a second result; and fusing the result 1 and the result 2 by using a pyramid fusion model, and outputting an enhanced underwater image. According to the method, the color naturalness, the overall contrast, the detail definition and the like of the underwater image can be effectively improved, and the quality of the underwater image is comprehensively improved. The method has important application value in the aspects of acquisition, storage, transmission, display, understanding and the like of underwater images / videos.

Description

Technical Field

[0001] The invention relates to underwater image enhancement technology, in particular to an adaptive underwater image enhancement method based on pyramid fusion, and belongs to the field of image processing. Background Art

[0002] Image is an important way to transmit information. It is widely used in all walks of life and plays an important role. In underwater applications, images also play an indispensable role. However, due to the complexity of the underwater environment, the acquisition and processing of underwater images face huge challenges. The current underwater imaging system cannot adapt to various underwater environments to ensure image quality. To solve this problem, improving the hardware conditions of the underwater imaging system is a direct method, but due to the high cost and limited effect. In contrast, improving the quality of underwater images through algorithm processing to meet the practical application of underwater images is a more comprehensive solution.

[0003] Underwater image enhancement is an image processing algorithm that aims to improve the contrast, color, clarity, and details of underwater images. In recent years, researchers have conducted research on underwater image enhancement technology from multiple perspectives, and proposed a series of algorithms based on traditional image enhancement technology, optical imaging models, Retinex models, deep learning, etc., which have successfully improved the quality of underwater images. However, various underwater image enhancement methods generally have problems such as color distortion, loss of local details, long running time, and low adaptability. For example, the optical model-based method often estimates underwater optical parameters for different scenes inaccurately or even too differently, and the deep learning-based method lacks high-quality underwater image reference data sets and other limitations. These problems make most existing underwater image enhancement methods unable to always achieve satisfactory visual enhancement effects when applied to low-quality images of real underwater scenes. Therefore, how to propose adaptive and effective enhancement models and implementation methods for the enhancement of real underwater scene images requires more in-depth research. Summary of the invention

[0004] In view of the above-mentioned shortcomings of the existing technology, this paper provides an adaptive underwater image enhancement method based on pyramid fusion.

[0005] In order to achieve the above-mentioned object of the invention, the present invention adopts the following technical solutions:

[0006] An adaptive underwater image enhancement method based on pyramid fusion mainly includes the following steps:

[0007] Step 1, compensating and balancing the blue and green channels of the input underwater image, and iteratively compensating the red channel using the balanced blue and green channels;

[0008] Step 2: Use a guided filter to reduce the noise of the color-corrected underwater image, and then globally stretch the pixel values ​​of its red, green, blue and brightness channels;

[0009] Step 3: for the stretched image, multi-scale filtering is used to extract and enhance its details to obtain result 1, and at the same time, histogram equalization with dynamic contrast limitation is performed to obtain result 2;

[0010] Step 4: Use the pyramid fusion model to fuse result 1 and result 2 and output the enhanced underwater image.

[0011] The present invention adopts the above technical scheme and has the following advantages: (1) Compared with the traditional white balance method, the present invention has good color correction ability for underwater images in various scenes and can effectively remove color casts of various colors. (2) The present invention dynamically selects the range of contrast stretching, and achieves a significant improvement in the overall contrast of the underwater image in the stretching of multiple channels, and the denoising process before stretching effectively suppresses the amplification of the noise component, making the stretched underwater image more realistic and natural. (3) The present invention improves the local details of the image from two angles. With the natural fusion effect of the pyramid fusion model, the results of these two detail enhancements complement each other, and the output underwater image details are clear and delicate. (4) The hardware requirements and time cost required to implement the present invention are relatively low, which is relatively easy to achieve in practical applications and has great practical application potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a general flow chart of the adaptive underwater image enhancement method based on pyramid fusion of the present invention;

[0013] Figure 2 It is a flow chart of the algorithm for color correction in the adaptive underwater image enhancement method based on pyramid fusion of the present invention;

[0014] Figure 3 It is a schematic diagram of a global stretching selection area in an adaptive underwater image enhancement method based on pyramid fusion of the present invention;

[0015] Figure 4 It is an algorithm flow chart of multi-scale detail enhancement in the adaptive underwater image enhancement method based on pyramid fusion of the present invention;

[0016] Figure 5 It is a fusion process diagram of a pyramid fusion model in the adaptive underwater image enhancement method based on pyramid fusion of the present invention;

[0017] Figure 6The following is a comparison of the enhancement results of the underwater test image by the present invention and the other three methods, where (a) is the input low-quality underwater image, (b) is the result of method 1, (c) is the result of method 2, (d) is the result of method 3, and (e) is the result after enhancement by the present invention:

[0018] Figure 7 The figure is a comparison of the effects of segmentation, key point detection, and saliency detection of a typical underwater image before and after enhancement by the present invention, wherein (a) is the processing result before enhancement, and (b) is the processing result after enhancement;

[0019] Figure 8 This is a diagram showing the defogging effect of a foggy image processed by the present invention. DETAILED DESCRIPTION

[0020] In order to make the present invention clearer, the present invention is described in more detail below in combination with specific embodiments and with reference to the accompanying drawings. It should be understood that the present invention can be implemented in various forms and is not limited by the accompanying drawings and the embodiments described below. The accompanying drawings and the embodiments described below are provided to enable the present invention to be more completely and accurately understood by those skilled in the art.

[0021] Figure 1 In this paper, an adaptive underwater image enhancement method based on pyramid fusion is proposed, which can be divided into the following four steps:

[0022] Step 1, compensating and balancing the blue and green channels of the input underwater image, and iteratively compensating the red channel using the balanced blue and green channels;

[0023] Step 2: Use a guided filter to reduce the noise of the color-corrected underwater image, and then globally stretch the pixel values ​​of its red, green, blue and brightness channels;

[0024] Step 3: for the stretched image, multi-scale filtering is used to extract and enhance its details to obtain result 1, and at the same time, histogram equalization with dynamic contrast limitation is performed to obtain result 2;

[0025] Step 4: Use the pyramid fusion model to fuse result 1 and result 2 and output the enhanced underwater image.

[0026] Specifically, in step 1, the algorithm flow is as follows: Figure 2 As shown. For the input underwater image I(x,y), decompose it into grayscale images of three channels: red, green and blue: I r (x,y),I g (x,y),I b (x, y). Before balancing the blue and green channels, compensation must be performed to avoid red color cast. First, in the green channel I g (x,y) and blue channel Ib Specifically, the channel with the larger average value among the two channels is selected as the compensation channel I. l (x, y), the channel with the smaller average value is the compensated channel I s (x,y), the compensation process is expressed as

[0027]

[0028] in, and Represents the average pixel value of the compensation channel and the compensated channel. The compensated blue and green channels are represented by express.

[0029] Furthermore, the balancing process first selects an area in the image that can represent the color of the entire image to estimate the overall pixel value of the blue-green channel. The local absolute error of the pixel value in this area is within the average range, and the conditions that need to be met are:

[0030]

[0031]

[0032] in, and Represents the pixel value and average value of the blue-green channel after compensation. M and N represent the height and width of the underwater image. It is called absolute error. The average pixel value of the area selected by the above formula is used to estimate the overall pixel value of the blue-green channel. and This further balances the blue and green channels, and this process is expressed as

[0033]

[0034] Further, in order to correct the red channel and prevent over-compensation of the red channel, which manifests itself as large areas of red artifacts or particles in the image. Specifically, the present invention uses the other two channels to constrain the red channel to prevent the pixel values ​​of certain areas of the red channel from being disproportionately high compared to the other two channels. The compensation increment of the red channel, that is, the target brightness of the red channel, is selected as the smaller value of the other two channels. It is expressed as

[0035]

[0036] And use it as the compensation increment and target intensity of the red channel, the iterative compensation process is expressed as

[0037]

[0038] in As the iteration process is continuously updated to ensure that the process meets the convergence condition, the stopping condition of the process is defined as

[0039] Δm<Δ th =10 -3 (7)

[0040] After the process converges quickly, the compensated red channel can be obtained, which is used together with the balanced blue-green channel as the final result of step 1. express.

[0041] Specifically, in step 2, in order to avoid amplifying the imaging noise generated by the harsh underwater environment, the artificial noise caused by the white balance operation, and the inherent quantization noise in the digital image, it is necessary to reduce the noise of the underwater image after color correction. The guided filter for denoising can be expressed as

[0042]

[0043] Where i and k represent the pixel positions, q i and I i denote the output image and the guide image respectively, ω i and |ω| represent the block at position i and the number of pixels in the block, respectively. The coefficient a k and b k Determined by the following formula

[0044]

[0045] where p i represents the input image, μ k and Respectively represent ω k The mean and variance within the block, ∈ is the regularization parameter. The guide image of the guided filter is the input image, and the block size is fixed to 10×10. Since the average block variance of the image represents the information contained in the image, the present invention makes First calculate the local variance of the 10×10 block of the image as

[0046]

[0047] The average block variance is calculated as

[0048]

[0049] Furthermore, the pixels are stretched globally in the three channels of red, green and blue synchronously. This process is expressed as

[0050]

[0051] in, represents the guided denoising filter mentioned above, and Represents the target maximum and minimum values ​​of the stretched image, set to 1 and 0. max and I min Represents the maximum and minimum values ​​of the image before stretching, but this maximum and minimum value are not the maximum and minimum values ​​in the narrow sense. In order to obtain a stable stretching result, the present invention estimates the mean of the brightest and darkest areas in the image. On the three channels of red, green and blue, the estimation strategy is as follows: the process of selecting the area is as follows Figure 3 As shown. First, the image to be stretched is divided into four rectangular blocks of equal size. Then the block with the smallest and largest average values ​​is selected from the four blocks and divided into four rectangular blocks. This process is repeated until the size of the rectangular block is less than the predetermined value. The specific exit condition is

[0052] M Ω N Ω <0.01×MN (13)

[0053] Among them, M Ω and N Ω Represent the height and width of the rectangular block respectively. Finally, the pixel averages of the two brightest and darkest rectangular blocks are used as the maximum and minimum estimates of the input image. The above operation obtains three groups of maximum and minimum values ​​of three channels. In order to avoid introducing color cast, they are unified into a group of maximum and minimum values. This process is expressed as

[0054]

[0055] Furthermore, in order to further enhance the effectiveness of the stretching process, the image is stretched globally in the CIELab color space. First, the stretched image is converted from the RGB color space to the CIELab color space, and the pixel values ​​are further stretched on the brightness channel, using the same strategy as above but without denoising. This process is represented as

[0056]

[0057] Convert the Lab image after the brightness channel is stretched back to the RGB image as the final result of step 2. express.

[0058] Specifically, in step 3, result 1 is obtained by extracting and enhancing details using a multi-scale filter. The algorithm flow chart for multi-scale detail extraction and enhancement is as follows: Figure 4 As shown in Figure 2, multi-scale is reflected in the use of Gaussian filter kernels of different sizes to extract details. For large-scale details, a Gaussian kernel with a diameter of d = 10 and a standard deviation of σ = 3 is used to directly extract the details of the stretched image. Convolve to get a smoothed version The original stretched image is subtracted from the smoothed version to obtain a large-scale detail image. This process is represented as

[0059]

[0060] For small-scale details, a Gaussian kernel with diameter d = 5 and standard deviation σ = 1.5 is used to calculate the input underwater image I c The restored version of (x,y) Convolve to get a smoothed version The original stretched image is subtracted from the smoothed version to obtain a large-scale detail image. This process is represented as

[0061]

[0062] This completes the extraction of large and small details and selectively superimposes them on the stretched image In order to effectively suppress noise, we use the result of formula (9) again and Get the detail gain factor

[0063]

[0064] The above equation means that blocks whose local variance is within three times of the average block variance will be enhanced in detail, and the larger their local variance, that is, the higher the noise component, the smaller the enhancement. The superposition process is expressed as

[0065]

[0066] The large-scale details have an additional gain of 0.5. That is result one.

[0067] Furthermore, the restoration process is based on the underwater optical imaging model, but the present invention only uses a simplified version because the present invention only needs to use its deblurring capability instead of the chromatic aberration correction capability. First, the dark channel is calculated. If The red channel is considered to have insignificant attenuation, and the underwater dark channel and underwater transmittance are estimated to be

[0068]

[0069]

[0070] like The red channel is considered to be significantly attenuated, and the underwater dark channel and underwater transmittance are estimated to be

[0071]

[0072]

[0073] Among them I r (y), I g (y) and I b (y) represents the three channels of the original input underwater image, and represents their respective average pixel values, and is the estimated underwater background light, expressed as

[0074]

[0075] where |p 10％ | and p 10％ are the number and position of the top 10% of pixel values ​​in the dark channel respectively. The above parameters are used to calculate the restored image based on the physical optical model. This process is represented as

[0076]

[0077] in same

[0078] Furthermore, the second result is obtained by contrast-limited adaptive histogram equalization on the brightness channel, and the limited contrast is dynamically determined by the local features. First, the image is converted to the CIELab color space to obtain the brightness channel. The present invention determines the upper limit of the contrast based on the information entropy of the local image after segmentation. The information entropy is calculated as follows:

[0079]

[0080] The pixels are converted to uint8 type with a value between 0 and 255. i Indicates the probability that the pixel value in the block is i. The larger the information entropy, the richer the information contained in this block of image. Therefore, the contrast upper limit set by the present invention is positively correlated with the information entropy. The histogram equalization process is expressed as

[0081] s k =T DCLAHE (r k ) (27)

[0082] where s k and r k Represents the grayscale of the underwater image brightness channel before and after equalization, T DCLAHE (·) represents a dynamic contrast-limited histogram equalization operation. The local details of the brightness channel are significantly improved, and then converted back to the RGB color space as result 2.

[0083] Specifically, in step 4, the main function of the pyramid fusion model is to fuse result 1 and result 2 to produce a stable output result. Figure 5 As shown. First, several features of result 1 and result 2 are extracted as weight maps, including chromaticity weight, saliency weight, exposure weight and gradient weight. The chromaticity weight is based on the fact that the positions with high saturation in the image should be fully enhanced. The process of calculating the chromaticity weight is expressed as

[0084]

[0085] Among them, S(x,y) represents the saturation channel image of result one or two in HSV color space, S max represents the maximum value of S(x,y), σ S = 0.3 represents the standard deviation. The saliency weights highlight the areas with high saliency, which is crucial in underwater images that are significantly attenuated. The process of calculating the saliency weights is expressed as

[0086]

[0087] Where L ωhc (x, y) represents the low-pass filtered version of the luminance channel in the CIELab color space, which uses a 5×5 Separable nucleus, and L μ The exposure weight plays a constant role in maintaining the appearance of local contrast. First, the mean value L of the brightness channel is calculated. μ and variance Then the exposure weight is calculated as

[0088]

[0089] In addition, gradient weights are introduced to increase the weight of details. The process is expressed as

[0090]

[0091] in Represents the first-order difference calculation of the horizontal and vertical gradients of the brightness channel.

[0092] Furthermore, for the two sets of weight graphs obtained: W chr1 (x,y),W sal1 (x,y),W exp1 (x,y),W gra1 (x,y) and W chr2 (x,y),W sal2 (x,y),W exp2 (x,y),W gra2(x,y), the process of normalization between corresponding weights is expressed as

[0093]

[0094] Where n∈{chr,sal,exp,gra} represents four weights, and k∈{1,2} represents result 1 and result 2. Then the sum of each set of weight graphs is calculated and normalized. This process is expressed as

[0095]

[0096]

[0097] Where δ = 10 -3 This is to avoid the normalization process causing some positions of the weight map to be 0 or ∞. The normalized weight map is used to calculate the Laplacian pyramid, and the Gaussian pyramid is calculated from the results 1 and 2. The layers are fused together to generate a natural-looking output image. This process is shown as

[0098]

[0099] Wherein, l represents the number of layers of the pyramid, and the maximum number of layers specified in the present invention is 5. and Represents the Laplacian pyramid of the image and Gaussian pyramid of normalized weight graph Represents the fusion result of each layer of the pyramid. Finally, the fusion result of the first layer That is, the enhanced output image

[0100] In order to verify the effectiveness of the method of the present invention, the present invention uses the standard test image set UIEB commonly used in the field of underwater image enhancement to conduct experiments. The UIEB set contains original low-quality underwater images of various underwater scenes, which are highly challenging in terms of color distortion, blur, low contrast, etc. In order to evaluate the underwater image enhancement method proposed in the present invention, three image enhancement algorithms are selected as comparison methods, specifically:

[0101] Method 1: The method proposed by Drews et al., reference "P.Drews, E.Nascimento, F.Moraes, S.Botelho, and M.Campos, "Transmission estimation in underwater single images," in Proceedings of the IEEE International Conference on Computer VisionWorkshops, (2013), pp.825–830."

[0102] Method 2: Azmi et al., reference “KZM Azmi, ASA Ghani, ZM Yusof, and Z. Ibrahim, “Natural-based underwater image color enhancement through fusion of swarm-intelligence algorithm,” Appl. Soft Comput. 85, 105810 (2019).”

[0103] Method 3: Islam et al., M.J. Islam, Y. Xia, and J. Sattar, “Fast underwater image enhancement for improved visual perception,” IEEE Robotics Autom. Lett. 5, 3227–3234 (2020).

[0104] The comparative experimental contents are as follows:

[0105] Table 1 lists the average SSIM index and average UCIQE index of different underwater image enhancement methods on the standard test image set UIEB. The SSIM index is used to evaluate the similarity of image structure information, while the UCIQE index combines chroma, saturation and contrast to evaluate the visual quality of the image. The larger the values ​​of these two indicators, the higher the objective quality of the image. It can be seen from Table 1 that the method of the present invention is significantly superior to other methods in terms of objective indicators, which fully proves the effectiveness of the present invention.

[0106] Figure 6 The effects of the method of the present invention and three other methods in processing underwater test images are compared. Figure 6The areas with obvious improvements have been marked with rectangular frames in the figure for easy observation and comparison. From the results, it can be observed that the method of the present invention not only improves the overall contrast of underwater images, but also makes the enhanced images present clearer, richer textures and details, while successfully suppressing adverse factors such as noise. Figure 6 The grayscale image in the figure cannot accurately show the effect of the method of the present invention in color correction, but in fact the method has also achieved remarkable results in color correction.

[0107] Figure 7 Taking the segmentation, key point detection and saliency detection of typical underwater images as examples, the effect of using the method of the present invention as a preprocessor on subsequent image analysis and understanding is demonstrated. Figure 7 By comparing the results in the figure, the following results can be observed: In the image segmentation task, the method of the present invention makes the outline of the segmentation result clearer and more realistic, and improves the accuracy of the overall structure of the image. In terms of image key point detection, the key points in the underwater image processed by the method of the present invention are more valuable, and the number of key points is significantly increased. In the image saliency detection task, the method of the present invention improves the accuracy and visual effect of the saliency detection results, and helps to better identify areas that people are more likely to pay attention to and remember. Therefore, the enhancement method of the present invention shows obvious advantages in subsequent applications such as image segmentation, key point detection and saliency detection, improves the quality and visibility of underwater images, and provides a reliable basis for further analysis and understanding.

[0108] Figure 8 The application effect of the present invention in other image enhancement fields, namely, defogging of foggy images, is demonstrated. The method of the present invention has achieved remarkable results in defogging, mainly reflected in the improvement of contrast and details. By applying the enhancement method of the present invention, the quality of images affected by foggy days can be adaptively improved to make them clearer, more contrasty, and richer in details. This proves that the method of the present invention has potential in the field of image defogging and provides an effective solution for improving images affected by foggy days.

[0109] In summary, the present invention is an effective underwater image enhancement method and has the potential to assist other image analysis and understanding tasks.

[0110] Table 1

[0111]

Claims

1. An adaptive underwater image enhancement method based on pyramid fusion, characterized in that The following steps are involved: Step 1, compensating and balancing the blue and green channels of the input underwater image, and iteratively compensating the red channel using the balanced blue and green channels; Step 2: Use a guided filter to reduce the noise of the color-corrected underwater image, and then globally stretch the pixel values ​​of its red, green, blue and brightness channels; Step 3: for the stretched image, multi-scale filtering is used to extract and enhance its details to obtain result 1, and at the same time, histogram equalization with dynamic contrast limitation is performed to obtain result 2; Step 4: Use the pyramid fusion model to fuse result 1 and result 2 and output the enhanced underwater image.

2. The method for adaptive underwater image enhancement based on pyramid fusion according to claim 1, characterized in that The blue-green channel compensation and balance are performed for the input underwater image with severe color deviation as described in step 1. The compensation process is expressed as follows: Among them, I l (x, y) represents the channel with the larger average value among the two channels, which is called the compensation channel. s (x, y) represents the channel with the smaller average value among the two channels, which is called the compensated channel. and Represents the average pixel value of the compensation channel and the compensated channel; Next, adaptive white balance is performed. The adaptive white balance is to select a specific area in the image that can represent the overall color of the image to estimate the pixel average value. The local absolute error of the pixel value in this area is within the average range. The conditions that need to be met are expressed as follows: in, and Represents the pixel value and average value of the blue-green channel after compensation. M and N represent the height and width of the underwater image. is called the absolute error, and the balancing process is expressed as in, and Represents the average pixel value of the above area.

3. The method for adaptive underwater image enhancement based on pyramid fusion according to claim 1, characterized in that The balanced blue-green channels described in step 1 iteratively compensate the red channel. The iterative process is expressed as in As the iteration process is continuously updated to ensure that the process meets the convergence condition, the stopping condition of the process is defined as Δm<Δ th =10 -3 After the process converges quickly, the compensated red channel can be obtained.

4. The method for adaptive underwater image enhancement based on pyramid fusion according to claim 1, characterized in that The determination of the regularization parameter of the guided denoising filter described in step 2 is expressed as in represents the local variance of a 10×10 block, represents the average block variance, ∈ is the regularization parameter used to determine whether the point is useless noise or useful detail.

5. The method for adaptive underwater image enhancement based on pyramid fusion according to claim 1, characterized in that The selection of the brightest and darkest areas and the determination of the maximum and minimum pixels described in step 2 first divide the image to be stretched into four rectangular blocks of equal size; then select the block with the smallest and largest average values ​​from the four blocks and divide it into four rectangular blocks; repeat this process until the size of the rectangular block is less than a predetermined value. The specific exit condition is M Ω N Ω <0.01×MN Among them, M Ω and N Ω Represent the height and width of the rectangular block respectively, M and N represent the height and width of the underwater image; finally, the pixel averages of the two brightest and darkest rectangular blocks are used as the maximum and minimum value estimates of the input image.

6. The method for adaptive underwater image enhancement based on pyramid fusion according to claim 1, characterized in that The multi-channel global contrast stretching described in step 2, the synchronous stretching process on the red, green and blue channels is expressed as Among them, c∈{r,g,b} represents the color channel, represents the guided denoising filter mentioned above, and Represents the target maximum and minimum values ​​of the stretched image, set to 1 and 0, I max and I min represents the maximum and minimum estimates of the image before stretching, which is obtained by averaging the maximum and minimum values ​​of three groups of three channels; similarly, the stretching process on the brightness channel is Where L S1 (x, y) is the brightness channel, L max and L min The method of obtaining I max and I min Same but no need to average, and Represents the target maximum and minimum values ​​of the stretched brightness image, also set to 1 and 0.

7. The method for adaptive underwater image enhancement based on pyramid fusion according to claim 1, characterized in that The multi-scale detail extraction method described in step 3, the extraction method of large and small scale details is expressed as in, is a Gaussian kernel with diameter d = 10 and standard deviation σ = 3, is a Gaussian kernel with diameter d = 5 and standard deviation σ = 1.5, represents the stretched underwater image, Represents the input underwater image I c The restored version of (x,y), Represents large-scale details, Represents small-scale details; after multi-scale extraction of details, they are selectively superimposed on the stretched underwater image, expressed as in, and are called local variance and average block variance respectively, which are calculated in step 2, β c (x,y) is called the detail gain coefficient and is used to selectively enhance parts with less details.

8. The method for adaptive underwater image enhancement based on pyramid fusion according to claim 1, characterized in that The underwater image restoration method described in step 3 first calculates the dark channel. The red channel is considered to have insignificant attenuation, and the underwater dark channel and underwater transmittance are estimated to be like The red channel is considered to be significantly attenuated, and the underwater dark channel and underwater transmittance are estimated to be Among them I r (y), I g (y) and I b (y) represents the three channels of the original input underwater image, and represents their respective average pixel values, and is the estimated underwater background light, expressed as where |p 10％ | and p 10％ are the number and position of the top 10% of pixel values ​​in the dark channel respectively; the restored image based on the physical optical model is calculated using the above parameters This process is represented as in same 9. The method for adaptive underwater image enhancement based on pyramid fusion according to claim 1, characterized in that The four weight map extraction processes described in step 4: Chroma weights are based on the fact that locations with high saturation in the image should be fully enhanced. The process of calculating chroma weights is expressed as Among them, S(x,y) represents the saturation channel image of result one or two in HSV color space, S max represents the maximum value of S(x,y), σ S = 0.3 represents the standard deviation; while the saliency weight highlights the area with high saliency, which is crucial in underwater images that are significantly attenuated. The process of calculating the saliency weight is expressed as in represents the low-pass filtered version of the luminance channel in the CIELab color space, which uses a 5×5 Separable nucleus, and L μ Represents the average value of the brightness channel; the exposure weight plays a constant role in maintaining the appearance of local contrast. First, the mean value L of the brightness channel is calculated μ and variance Then the exposure weight is calculated as In addition, gradient weights are introduced to increase the weight of details. The process is expressed as in Represents the first-order difference calculation of the horizontal and vertical gradients of the brightness channel.

10. The method for adaptive underwater image enhancement based on pyramid fusion according to claim 1, characterized in that The weight map normalization and fusion process described in step 4, the normalization process is expressed as Where n∈{chr,sal,exp,gra} represents four weights, k∈{1,2} represents the two results of step 3, δ=10 -3 This is to avoid that some positions of the weight map are 0 or ∞ during the normalization process; the fusion process is expressed as Wherein, l represents the number of layers of the pyramid, and the maximum number of layers specified in the present invention is 5, and and Represents the Laplacian pyramid of the image and Gaussian pyramid of normalized weight graph Represents the fusion result of each layer of the pyramid, and the first layer fusion result That is, the enhanced output image

Images