Image dynamic range compression method based on deep learning and joint bilateral filtering

Abstract

The application discloses a kind of based on deep learning and joint bilateral filtering's image dynamic range compression method, including steps: collecting original high bit image data, constructs data set, and preliminary processing is carried out to data set, obtains linear RGB graph and Y channel graph;Combined with the skip connection structure of UNet and the efficient light unit of ShuffleNet constructs deep learning network, utilizes deep learning network to linear RGB graph and Y channel graph processing, obtains luminance gain graph;With the help of joint bilateral filtering upsampling operation, luminance gain graph and original high bit image are processed, and original resolution luminance gain graph is obtained;Original high bit image is carried out inverse gamma operation, and then divided by original resolution luminance gain graph is carried out gamma operation to obtain compressed 8bit image.The application realizes a dynamic range adjustment algorithm, from high bit to low bit mapping process, beneficial to restore highlight details, brighten dark details, and maintain good image contrast.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to an image dynamic range compression method based on deep learning and joint bilateral filtering. BACKGROUND

[0002] In the field of image style imaging technology, dynamic range adjustment is a key link to realize high-quality image display. The traditional tone mapping algorithm adjusts the brightness curve of the image to compress the high dynamic range image to the range that can be presented by the display device, and is widely used in image and video processing. However, this algorithm has limitations in practical application. It relies on a fixed mathematical transformation model and is difficult to adaptively process complex light distribution of images in different scenes, which can easily lead to loss of highlight details or blurring of dark details, and cannot fully meet the user's requirements for image detail restoration and visual effects.

[0003] With the development of mobile devices and photography technology, mobile phones, single-lens reflex cameras and other devices have become the mainstream tools for image acquisition. During the shooting process, users often encounter high dynamic range scenes, such as backlight environment, scenes with strong light and shadow, etc., which puts higher requirements on the dynamic range processing of the device. The imaging system of the existing device is difficult to achieve detail preservation in high light area, detail enhancement in dark area and good image contrast at the same time when facing such scenes. Therefore, there is an urgent need for a more efficient and intelligent image dynamic range compression technology to improve the visual quality and display effect of the image and meet the needs of users in diversified shooting scenes. SUMMARY

[0004] In order to solve the above technical problems, in view of the deficiencies in the background art, the purpose of the present application is to provide an image dynamic range compression method based on deep learning and joint bilateral filtering. The method is applied to the shooting path of mobile phones and single-lens reflex cameras. Given an image with high dynamic range and high resolution, the image is first down-sampled, then the down-sampled high dynamic range image is subjected to deep learning, and after obtaining the result, the original input image, the down-sampled input image and the output image obtained by the network after down-sampling are subjected to joint bilateral filtering up-sampling, and finally a good dynamic range compressed image is obtained, which can restore high light details, brighten dark details and maintain good image contrast. Finally, a displayable high dynamic range image is obtained, solving the problem of loss of highlight details or blurring of dark details.

[0005] The technical solution adopted by the present application is as follows:

[0006] S1) Collecting original high-bit image data, constructing a data set, and preliminarily processing the data set to obtain a linear RGB image and a Y channel image;

[0007] S2) Construct a deep learning network by combining the skip connection structure of UNet and the efficient lightweight unit of ShuffleNet, and process the linear RGB image and the Y channel image using the deep learning network to obtain a luminance gain image;

[0008] S3) Process the luminance gain image and the original high-bit image by means of a joint bilateral filter upsampling operation to obtain an original resolution luminance gain image;

[0009] S4) Perform an inverse gamma operation on the original high-bit image, divide it by the original resolution luminance gain image, and then perform a gamma operation to obtain a compressed 8-bit image.

[0010] Preferably, in step S1), the data set is preliminarily processed to obtain a linear RGB image and a Y channel image, and the specific process is as follows:

[0011] Sa1) Down-sampling operation: the original high-bit image in the data set is down-sampled by bilinear interpolation, and the new pixel value of the target pixel is calculated by linearly weighting and averaging the four nearest neighbor pixels around the target pixel, so that the resolution of the down-sampled image is 512x512, thereby reducing the resolution and the computational load of subsequent processing;

[0012] Sa2) Inverse gamma operation and color gamut conversion operation: the down-sampled image is subjected to an inverse gamma operation pixel by pixel, and the inverse gamma operation adopts the inverse function of the gamma operation formula , which converts the non-linear domain image to a linear domain image to obtain a linear RGB image, i.e. , wherein is 2.2, is the image after inverse gamma operation, is the image before inverse gamma operation; at the same time, the down-sampled image is subjected to an RGB to Y color gamut conversion operation, i.e. an image containing an RGB color space is converted into a Y channel image containing luminance information, and the formula is Y=0.2126R+0.0722B+0.7125G, wherein Y is the luminance value obtained after conversion, R is the red channel value of the pixel in the image in the RGB color space, B is the green channel value of the pixel in the image in the RGB color space, and G is the blue channel value of the pixel in the image in the RGB color space.

[0013] Preferably, in step S2), a deep learning network is constructed by combining the skip connection structure of UNet and the efficient lightweight unit of ShuffleNet, and the linear RGB image and the Y channel image are processed using the deep learning network to obtain a luminance gain image, and the process is as follows:

[0014] Sb1) constructing a deep learning network: adopting a deep learning network structure combined with a skip connection structure of UNet and an efficient lightweight unit of ShuffleNet, by an Encoder part, a Decoder part and a Bottleneck part; wherein the Encoder part adopts an efficient lightweight unit of ShuffleNetV2 based on ShuffleNet, introduces a channel shuffle operation processing, gradually reduces the resolution of the image through a series of convolution operations, while increasing the number of feature channels, to realize the extraction of multi-level features; the Decoder part adopts the skip connection structure of UNet, gradually restores the spatial resolution of the image, at the same time, through the skip connection, the multi-level features of the Encoder part are fused with the features processed by the Decoder part, to restore the spatial detail information of the image; the Bottleneck part uses the depth separable convolution structure of ShuffleNet, to obtain a compressed and colored nonlinear image through depth separable convolution processing, to reduce the number of parameters and the amount of calculation of convolution operation;

[0015] Sb2) input of the deep learning network: the linear RGB image obtained by inverse gamma and the Y channel Figure One of the original high-bit image are taken as the input of the deep learning network;

[0016] Sb3) output and processing of the deep learning network: outputting the y nonlinear domain image after dynamic range adjustment through the deep learning network, and then performing inverse gamma operation to obtain the y linear domain image, at the same time, performing inverse gamma operation on the linear RGB image to obtain the Y linear domain image, and obtaining the luminance gain image with a resolution of 512x512 by dividing the Y linear domain image by the y linear domain image.

[0017] Preferably, in step S3), the luminance gain image and the original high-bit image are processed by means of joint bilateral filtering upsampling operation to obtain the original resolution luminance gain image, and the specific operation steps include:

[0018] by taking the low-resolution luminance gain image and the high-resolution original high-bit image with the resolution size before downsampling as inputs, finding the 5x5 size local window region corresponding to each pixel point in the original high-bit image in the luminance gain image , and through the way of bilinear interpolation, the luminance gain image is interpolated to the high-resolution grid, and for each pixel in the original high-bit image , its value is obtained through the luminance gain image Central neighborhood pixels weighted average of the central neighborhood pixels , the calculation formula is:

[0019]

[0020] In the formula is a normalization factor, is a joint weight;

[0021] wherein, the calculation formula of the joint weight is:

[0022]

[0023] In the formula is the standard deviation of the control space weight; is the standard deviation of the control reference image similarity; is the pixel and the pixel spatial distance square; is the reference image difference of the pixel and the pixel ; is an exponential function.

[0024] Preferably, in step S4), the original high bit image is subjected to inverse gamma operation, and then divided by the original resolution luminance gain map to obtain the compressed 8bit image through gamma operation, and the specific operation steps include:

[0025] The original high bit image is subjected to inverse gamma operation to obtain a linear domain image, and then divided by the original resolution luminance gain map to restore the image, thereby obtaining an original resolution image in the linear domain after dynamic range adjustment, and then subjected to gamma operation to obtain a compressed 8bit image.

[0026] Compared with the prior art, the image dynamic range compression method based on deep learning and joint bilateral filtering has the following advantages:

[0027] The dynamic range adjustment algorithm realizes the mapping process from high bit to low bit. In terms of effect, compared with the traditional algorithm, the dynamic range visualization effect is better, the image dark area can be effectively improved, and the highlight area of the image can be restored. In terms of algorithm efficiency, the image is changed into a small image through a simple deep learning network, and then up-sampling is performed through joint bilateral filtering, so that the operation performance efficiency can be effectively improved, and there is no light spot defect after up-sampling processing. In the image algorithm processing, it is a high-efficiency and excellent image quality algorithm. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 The flow framework diagram of the image dynamic range compression method based on deep learning and joint bilateral filtering proposed in the application is as shown in

[0029] Figure 2 The deep learning network framework diagram proposed in the application is as shown in

[0030] Figure 3 The contrast diagram of the indoor original high-bit image and the 8-bit image of the application is as shown in the figure, wherein a is the indoor original high-bit image, and b is the contrast diagram of the indoor 8-bit image;

[0031] Figure 4 The contrast diagram of the outdoor original high-bit image and the 8-bit image of the application is as shown in the figure, wherein c is the outdoor original high-bit image, and d is the contrast diagram of the outdoor 8-bit image. DETAILED DESCRIPTION

[0032] The technical solutions in the embodiments of the application will be further described in detail below with reference to the accompanying drawings of the embodiments of the application. It should be explained that the described embodiments are only some of the embodiments of the application, but not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the application.

[0033] In order to make the application purposes, technical solutions and advantages of the application more clear, the embodiments of the application will be further described in detail below with reference to the accompanying drawings of the application. In order to more clearly understand the above purposes, features and advantages of the application, the advantages of the application will be further described by comparing the embodiments with reference to the accompanying drawings and specific embodiments.

[0034] The technical solutions of the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0035] The image dynamic range compression method based on deep learning and joint bilateral filtering proposed in the application has the flow framework diagram as shown in Figure 1 The method comprises the following steps:

[0036] S1) Collecting original high-bit image data, constructing a data set, and performing preliminary processing on the data set to obtain a linear RGB image and a Y channel image;

[0037] Specifically, the data set is preliminarily processed in step S1) to obtain a linear RGB image and a Y channel image, and the specific process is as follows:

[0038] Sa1) downsampling operation: the original high-bit image in the input data set has a high resolution, and the image is subjected to a downsampling operation in a bilinear interpolation manner, four nearest neighbor pixels around the target pixel of the image are selected, and the new pixel value of the image target is calculated by linearly weighting and averaging the four nearest neighbor pixels, so that the resolution of the image after the downsampling operation is 512*512, so as to reduce the resolution and the calculation amount of subsequent processing;

[0039] Sa2) inverse gamma operation and color gamut conversion operation: the image after the downsampling processing is subjected to an inverse gamma operation pixel by pixel, and a color gamut conversion operation is performed on the image after the inverse gamma operation, that is, the image is converted from the RGB color space to the Y channel space, and the formula is Y=0.2126R+0.0722B+0.7125G, wherein Y is the luminance value obtained after conversion, R is the red channel value of the pixel point in the image in the RGB color space, B is the green channel value of the pixel point in the image in the RGB color space, and G is the blue channel value of the pixel point in the image in the RGB color space (wherein the coefficients 0.2126, 0.0722 and 0.7125 in the formula are obtained based on experimental data of the sensitivity of the human eye to different colors of light). is 2.2, is the image after the gamma operation, is the image before the gamma operation) is a process of nonlinear luminance transformation of the image, and the purpose is to match the nonlinear response characteristics of the human eye to the intensity of light; and the inverse gamma operation is the inverse operation of the gamma operation, which converts the nonlinear domain image into a linear domain image, and prepares for subsequent processing; the inverse gamma operation adopts the inverse function of the gamma operation formula, to obtain a linear RGB image, that is , wherein is 2.2, is the image after the inverse gamma operation, is the image before the inverse gamma operation; at the same time, the image after the downsampling processing is subjected to an RGB to Y color gamut conversion operation, that is, the image containing the RGB color space is converted into a Y channel image containing luminance information, and the formula is Y=0.2126R+0.0722B+0.7125G, wherein Y is the luminance value obtained after conversion, R is the red channel value of the pixel point in the image in the RGB color space, B is the green channel value of the pixel point in the image in the RGB color space, and G is the blue channel value of the pixel point in the image in the RGB color space (wherein the coefficients 0.2126, 0.0722 and 0.7125 in the formula are obtained based on experimental data of the sensitivity of the human eye to different colors of light).

[0040] S2) a deep learning network is constructed by combining the skip connection structure of UNet and the efficient lightweight unit of ShuffleNet, and the linear RGB image and the Y channel image are processed by using the deep learning network to obtain a luminance gain image;

[0041] Specifically, in step S2), a deep learning network is constructed by combining the skip connection structure of UNet and the efficient lightweight unit of ShuffleNet, and the linear RGB image and the Y channel image are processed by using the deep learning network to obtain a luminance gain image, and the deep learning network framework diagram is as follows​Figure 2 The process is as follows:

[0042] Sb1) Deep learning network structure: a deep learning network structure is constructed by combining the jump connection structure of UNet and the efficient lightweight unit of ShuffleNet, which includes an Encoder part, a Decoder part and a Bottleneck part; the efficient lightweight unit of ShuffleNetV2 based on ShuffleNet is used in the Encoder part, a channel shuffling operation is introduced for processing, and a series of convolution operations are used to gradually reduce the image resolution while increasing the number of feature channels, thereby achieving multi-level feature extraction; the jump connection structure of UNet is used in the Decoder part to gradually restore the resolution of the linear image, and at the same time, the multi-level features of the Encoder part are fused with the features processed by the Decoder part through the jump connection to restore the spatial detail information of the image, so that the network can reconstruct an image with rich details; the Bottleneck part uses the depth separable convolution structure of ShuffleNet to obtain a Y graph through depth separable convolution processing, thereby reducing the number of parameters and the amount of calculation of convolution operation;

[0043] The advantages of the deep learning network structure are: ensuring speed and performance, which is conducive to use in mobile terminals;

[0044] ① Progressive down-sampling and feature abstraction (Encoder part)

[0045] Optimization of hierarchical compression ratio:

[0046] Through 4-stage down-sampling (512→16 resolution), 16x spatial compression is achieved, and ShuffleNet Unit with Stride=2 is used in each stage instead of traditional pooling layer for down-sampling, which avoids information loss of traditional pooling and realizes multi-level feature extraction, as follows:

[0047] Initially, Conv3x3 convolution operation and MaxPool operation are used to quickly reduce the resolution (512→128) while retaining 24-channel basic features;

[0048] In Encoder3, the deep ShuffleNet Unit enhances 32x32 scale feature extraction through 3 times of repetition (Repeat=3), and the number of channels is 232, which is close to the number of channels of VGG16 in the same stage (256), but the number of parameters is reduced by 37%;

[0049] In the Encoder2 (64x64x116) stage, through group convolution and channel rearrangement, the feature reuse rate is improved by about 2.1 times through actual comparison tests;

[0050] ②Bottleneck's channel expansion design (Bottleneck section)

[0051] Efficiency of deep separable convolution:

[0052] In the process of generating 16x16x928 feature maps, the standard convolution calculation is 928x464x3x3=3.9MFLOPs, and the actual deep separable convolution is 464x3x3+928x464x1x1=0.43MFLOPs. By splitting the standard convolution into deep separable convolution, the calculation is reduced by 89%; and using non-downsampling operation (Stride=1) to preserve the integrity of spatial information;

[0053] ③Symmetric up-sampling and jump connection (Decoder section)

[0054] Layer-by-layer feature fusion mechanism:

[0055] Size matching: After each UpConv2x2, the corresponding Encoder feature is spliced (such as Decoder4's 32x32x928 and Encoder4's 32x32x464), and the channel number is unified through 1x1 convolution (232) to avoid information overload;

[0056] Decoder application of ShuffleNet Unit: At the 32x32 scale, the 1x1→3x3→1x1 structure of ShuffleNet Unit (parameter amount is only 1 / 5 of the standard residual block) is applied to realize efficient feature fusion, and the inference speed is 2.3 times faster than traditional UNet;

[0057] ④End-to-end design advantages

[0058] Input-output consistency:

[0059] Directly output 512x512x1 through Conv1x1 convolution, and save redundant up-sampling (such as 32 times up-sampling of FCN);

[0060] Mobile adaptation:

[0061] The whole network has only 3.8M parameters, which can be quantized to INT8 (accuracy loss <0.5%), and the actual power consumption of Snapdragon865 chip is <1.2W, which meets the real-time requirement;

[0062] Sb2) Input of deep learning network: linear RGB image and Y channel image (total of 4 channels) obtained by inverse gamma are used as input of deep learning network;

[0063] Sb3) output processing of the deep learning network: output the y nonlinear domain image after dynamic range adjustment after processing by the deep learning network, and obtain the y linear domain image by performing inverse gamma operation thereon, and obtain the Y linear domain image by performing inverse gamma operation on the linear RGB image, and obtain the 512x512 luminance gain image by dividing the Y linear domain image by the y linear domain image.

[0064] S3) processing the luminance gain image and the original high-bit image by means of the joint bilateral filter upsampling operation to obtain the original resolution luminance gain image;

[0065] Specifically, in step S3), the luminance gain image and the original high-bit image are processed by means of the joint bilateral filter upsampling operation to obtain the original resolution luminance gain image, and the specific operation steps include:

[0066] The joint bilateral filter upsampling is an image super-resolution technology based on edge preservation, which performs high-quality upsampling on a low-resolution image (such as a depth map or a semantic label) by combining the geometric structure of a high-resolution image (such as color information of an RGB or depth map), and the core idea is to use the weighting mechanism of bilateral filtering to preserve edge details while enlarging the image;

[0067] By taking the low-resolution luminance gain image and the high-resolution original high-bit image of the resolution size before downsampling as inputs, finding the 5x5 size local window area corresponding to the pixel point in the original high-bit image in the luminance gain image , and interpolating the luminance gain image to the high-resolution grid by means of bilinear interpolation, for each pixel in the original high-bit image , its value is obtained by weighted average of the neighborhood pixels in the luminance gain image , to obtain the original resolution luminance gain image , and the calculation formula is:

[0068]

[0069] In the formula, is a normalization factor, is a joint weight;

[0070] The joint weight is determined by the spatial distance and the reference image intensity difference, which determines the pixel in the luminance gain image The contribution size of the pixel of the original high-bit image The spatial distance The intensity difference of the color and the brightness of the pixel of the original high-bit image The spatial distance The intensity difference of the color and the brightness of the pixel of the original high-bit image The spatial distance The intensity difference of the color and the brightness of the pixel of the original high-bit image The spatial distance The intensity difference of the color and the brightness of the pixel of the original high-bit image The spatial distance The intensity difference of the color and the brightness of the pixel of the original high-bit image The spatial distance The intensity difference of the color and the brightness of the pixel of the original high-bit image

[0071] The spatial distance The intensity difference of the color and the brightness of the pixel of the original high-bit image The spatial distance

[0072] The spatial distance The spatial distance The spatial distance The spatial distance The spatial distance The spatial distance The spatial distance The spatial distance The spatial distance The spatial distance The spatial distance The spatial distance The spatial distance

[0073] S4) inverse gamma operation on the original high-bit image, and then divide by the original resolution luminance gain map to obtain the compressed 8-bit image through gamma operation;

[0074] Specifically, in step S4), two groups of original high-bit images (the first group is an indoor image, and the second group is an outdoor image) are used, inverse gamma operation is performed on the original high-bit image, and then the original resolution luminance gain map is divided to obtain the compressed 8-bit image through gamma operation. The comparison chart of the obtained original high-bit image and 8-bit image is shown in FIGS. 1 and 2, and the specific operation steps include: Figure 3 Figure 4 The original high-bit image is subjected to inverse gamma operation to obtain a linear domain image, and then the original resolution luminance gain map is divided to restore the image, thereby obtaining the original resolution image after dynamic range adjustment in the linear domain, and then gamma operation is performed to obtain the compressed 8-bit image.

[0075] The original high-bit image is subjected to inverse gamma operation to obtain a linear domain image, and then the original resolution luminance gain map is divided to restore the image, thereby obtaining the original resolution image after dynamic range adjustment in the linear domain, and then gamma operation is performed to obtain the compressed 8-bit image.

[0076] The content not described in detail in the specification belongs to the prior art known to those skilled in the art.

[0077] Finally, it should be noted that the above specific embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the protection scope thereof. Although the present application has been described in detail in conjunction with the preferred embodiments, it should be understood by those skilled in the art that, without departing from the core idea of the present application, modifications, replacements or equivalent changes can be made to the technical content thereof, and all should be considered to fall within the protection scope of the present application, and the protection scope of the present application should be limited by the content defined in the claims.

Claims

1. A deep learning and joint bilateral filtering based image dynamic range compression method, characterized in that, The method comprises the following steps: S1) collecting original high-bit image data, constructing a data set, and performing preliminary processing on the data set to obtain a linear RGB graph and a Y channel graph; S2) combining the skip connection structure of UNet and the efficient lightweight unit of ShuffleNet to construct a deep learning network, and using the deep learning network to process the linear RGB graph and the Y channel graph to obtain a brightness gain graph; Sb1) constructing a deep learning network: adopting an Encoder part, a Bottleneck part and a Decoder part, combining the skip connection structure of UNet and the efficient lightweight unit of ShuffleNet to construct a deep learning network structure; wherein the Encoder part adopts a ShuffleNetV2 efficient lightweight unit based on ShuffleNet, introduces a channel shuffling operation for processing, gradually reduces the resolution of the image through a series of convolution operations, and increases the number of feature channels at the same time, realizing the extraction of multi-level features; the Bottleneck part uses the depth separable convolution structure of ShuffleNet, and obtains a compressed and colored nonlinear image through depth separable convolution processing, reducing the number of parameters and the amount of calculation of convolution operation; the Decoder part adopts the skip connection structure of UNet, gradually restores the spatial resolution of the image, and at the same time, the multi-level features of the Encoder part and the features processed by the Decoder part are fused through the skip connection to restore the spatial detail information of the image; Sb2) input of the deep learning network: the linear RGB graph and the Y channel graph obtained through inverse gamma are used as the input of the deep learning network; Sb3) output and processing of the deep learning network: the y nonlinear domain graph after dynamic range adjustment is output through the deep learning network, and then inverse gamma operation is performed to obtain the y linear domain graph; at the same time, inverse gamma operation is performed on the linear RGB graph to obtain the Y linear domain graph, and the brightness gain graph with a resolution of 512x512 is obtained by dividing the Y linear domain graph by the y linear domain graph; S3) processing the brightness gain graph and the original high-bit image by means of joint bilateral filtering upsampling operation to obtain an original resolution brightness gain graph; S4) performing inverse gamma operation on the original high-bit image, then dividing by the original resolution brightness gain graph and performing gamma operation to obtain a compressed 8-bit image.

2. The image dynamic range compression method based on deep learning and joint bilateral filtering according to claim 1, characterized in that, In step S1), the data set is preliminarily processed to obtain a linear RGB graph and a Y channel graph, and the specific process is as follows: Sa1) downsampling operation: the original high-bit image in the data set is input, and the image is subjected to downsampling operation in a bilinear interpolation manner, four nearest neighbor pixels around the target pixel of the image are selected, and the new pixel value of the target pixel is calculated by linearly weighting and averaging the four nearest neighbor pixels, so that the resolution of the image after downsampling operation is 512x512, so as to reduce the resolution and the calculation amount of subsequent processing; Sa2) inverse gamma operation and gamut conversion operation: inverse gamma operation is performed on the image after down-sampling processing, pixel by pixel, and the inverse gamma operation adopts the gamma operation The inverse function of the formula converts the nonlinear domain image into the linear domain image to obtain a linear RGB image, i.e. , wherein is 2.2, is the image after inverse gamma operation, is the image before inverse gamma operation; meanwhile, the RGB-to-Y gamut conversion operation is performed on the image after down-sampling processing, i.e. the image containing the RGB color space is converted into the Y channel image containing luminance information, and the formula is Y=0.2126R+0.0722B+0.7125G, wherein Y is the luminance value obtained after conversion, R is the red channel value of a pixel in the image in the RGB color space, B is the green channel value of a pixel in the image in the RGB color space, and G is the blue channel value of a pixel in the image in the RGB color space.

3. The image dynamic range compression method based on deep learning and joint bilateral filtering according to claim 1, characterized in that, In step S3, the luminance gain map and the original high-bit image are processed by means of a joint bilateral up-sampling operation to obtain an original resolution luminance gain map, and the specific operation steps include: by combining a low resolution luma gain map and a high resolution raw high bit image of pre-downsampled front resolution size as input, find in the luma gain map the corresponding 5x5 size local window region to the pixel in the raw high bit image , interpolate the luma gain map to the high resolution grid by bilinear interpolation, for each pixel in the raw high bit image , its value is obtained by weighted average of the neighborhood pixels in the luma gain map , the raw resolution luma gain map is calculated as ； wherein is a normalization factor, is a joint weight; wherein the joint weight is calculated as The calculation formula is: ； wherein is a standard deviation of spatial weights; is a standard deviation of reference image similarity; is a pixel and a pixel a square of spatial distance; is a pixel and a pixel reference image difference of pixels is an exponential function.

4. The image dynamic range compression method based on deep learning and joint bilateral filtering according to claim 1, characterized in that, In step S4, the original high-bit image is subjected to inverse gamma operation, and then divided by the original resolution luminance gain map to obtain a compressed 8-bit image through gamma operation, and the specific operation steps include: The original high-bit image is subjected to inverse gamma operation to obtain a linear domain image, and then divided by the original resolution luminance gain map to restore the image, thereby obtaining a linear domain original resolution image after dynamic range adjustment, and then subjected to gamma operation to obtain a compressed 8-bit image.

Images