A method to solve the SNR drop problem after HDR image fusion

Abstract

The present application provides a kind of method for solving the problem of SNR Drop after HDR image fusion, comprising: S1.data acquisition;S2.HDR image label production;S3.select training pair combination;S4.data augmentation;S5.data clipping;S6.design convolutional neural network;S7.train convolutional neural network;S8.model inference post-processing and visualization.The HDR fusion algorithm proposed in this method only uses two frames, while the traditional algorithm improves the SNR Drop phenomenon above the synthesis of multiple frames in succession, and the ghost problem in synthesis is prone to occur, the present application has relatively low requirements for image sensor, and can avoid the ghost problem in synthesis as much as possible;The proposed HDR fusion algorithm does not reduce the long-short frame exposure ratio of the image sensor, does not lose the dynamic range of the scene, and keeps the details of the bright and dark areas of the scene well.

Description

Technical Field

[0001] This invention belongs to the field of intelligent surveillance video processing technology, and specifically relates to a method for solving the SNRDrop problem after HDR image fusion. Background Technology

[0002] If a traditional image sensor is used to capture a high dynamic range (HDR) scene, either bright areas will be overexposed, losing a lot of detail, or dark areas will be underexposed, making it difficult to discern details. In order to record every detail of a HDR scene, people use a regular sensor to obtain several frames of fixed scene images with different exposures, and then use HDR algorithms to synthesize a single HDR image.

[0003] Traditional HDR algorithms for real image sensors fuse two or three frames of images with different exposures. For scenes with a large dynamic range, the brightness difference between the two or three frames of images with different exposures is large, and the exposure difference between the two frames is even more obvious, resulting in a large difference between the bright and dark areas. The fused image exhibits the SNR Drop phenomenon at the boundary between bright and dark areas.

[0004] To mitigate the SNR Drop phenomenon, traditional HDR algorithms typically reduce the difference between bright and dark boundaries by combining multiple frames with different exposures or by lowering the exposure ratio.

[0005] However, using multi-frame image synthesis with different exposures is difficult to use in real time due to the limitations of image sensor computing resources, and the probability of synthesized ghosting is relatively high due to the introduction of multi-frame synthesis; reducing the exposure ratio, although reducing the difference between bright and dark boundaries, loses the dynamic range of the scene, and the details of bright and dark areas cannot be preserved at the same time.

[0006] In addition, commonly used technical terms in existing technologies include:

[0007] HDR stands for High Dynamic Range. Dynamic range refers to the ratio of the brightness of the brightest object to the darkest object in a scene. The larger the dynamic range, the richer the tonal range it can represent. Real-world scenes achieve a dynamic range of... The dynamic range that the human visual system can perceive is approximately... The dynamic range of a typical image sensor is approximately... .

[0008] SNR Drop: A decrease in signal-to-noise ratio. SNR drop is a phenomenon caused by the non-linear SNR curve produced by the multi-exposure HDR algorithm. This problem is common in bright areas where multiple images are merged. When the SNR drop is large, obvious edge stitching lines will appear.

[0009] Exposure ratio: Typically used in HDR scenarios, it controls the exposure time of an image sensor continuously exposing images.

[0010] Convolutional Neural Networks (CNNs): A CNN is a deep neural network with a convolutional structure. Images can be directly used as input to the network, avoiding the complex feature extraction and data reconstruction processes in traditional algorithms. It has great advantages in the processing of two-dimensional images, such as the ability of the network to extract image features including color, texture, shape and topological structure. It is also robust and efficient in processing two-dimensional images.

[0011] Bayer format: Bayer format is the raw image data inside the image sensor. Generally, raw data consists of 1 / 2 G, 1 / 4 R, and 1 / 4 B. The image is processed through a series of image processing steps to become the color image that people see. Summary of the Invention

[0012] To address the aforementioned issues, the purpose of this application is to design an HDR synthesis algorithm using a convolutional neural network. This algorithm fuses two frames of images, which places lower demands on the computational resources of the image sensor, avoids the ghosting problem introduced by multi-frame synthesis, and can better fuse details in bright and dark areas without losing the dynamic range of the scene, thus solving the SNR Drop phenomenon.

[0013] Specifically, the present invention provides a method for solving the SNR Drop problem after HDR image fusion, the method comprising the following steps:

[0014] S1. Data Collection:

[0015] Different exposure data were collected in various high dynamic range static scenes. Two frames of raw data, one long and one short, were collected at once using an image sensor. Raw data was collected in high dynamic range static scenes with different appropriate exposure times. Ten frames were collected for each scene to ensure that the synthesized HDR image can cover the dynamic range of the scene.

[0016] S2. HDR Image Tagging Creation:

[0017] HDR image labels for each high dynamic range still scene are generated based on the traditional weighted fusion algorithm.

[0018] S3. Select training pair combinations:

[0019] S3.1, each high dynamic range still scene has 10 frames of raw data, which are combined in pairs. The HDR image synthesized by each combination is calculated according to Equation (1). The curve mapping result of the HDR image and the HDR image label of each combination are calculated according to Equation (2). ,

[0020] Equation (2)

[0021] in For HDR images, The constant is 5000;

[0022] S3.2, the curve mapping results corresponding to the HDR images and the curve mapping results corresponding to the HDR image labels for each combination are calculated according to equation (3). ,choose The largest combination is used as the training input data, which, together with the HDR image labels, forms the training pair combination corresponding to the high dynamic range still scene:

[0023] Equation (3)

[0024] in and These represent the width and height of the HDR image, respectively. and These are the curve mapping results corresponding to HDR images and the curve mapping results corresponding to HDR image labels, respectively.

[0025] S4. Data Augmentation:

[0026] The raw data is in Bayer format, assuming it is rggb Bayer. The r and b channels of long and short frames are multiplied by 1.0, 1.2, 1.4, 1.6, 1.8, and 2.0 and then saved to increase the diversity of the raw data. The traditional weighted fusion algorithm in step S2 is then applied to the increased data to obtain the HDR image label of the increased data.

[0027] S5. Data cropping: Cropping the original raw data and HDR image labels of the training pair combination into... Size, increasing data diversity;

[0028] S6. Designing Convolutional Neural Networks:

[0029] S6.1, Design the network input; the raw data is in Bayer format, assuming it is... The resolution is obtained by extracting the raw data into the rggb Bayer format, which extracts the four channels (r, g, g, b) sequentially to obtain the resolution. Without losing original information, the input resolution is reduced to decrease the network computational load. The input is a superposition of two frames of data, resulting in a final network input resolution of [resolution value missing]. ;

[0030] S6.2, Design the input data preprocessing: extract the two frames of raw data into four channels and then normalize the data. and Divide by the corresponding exposure time respectively and As network input;

[0031] S6.3, Design the feature extraction network. The basic feature extraction first uses a convolution with a kernel of 3x3, a stride of 1, and a pad of 1 and Leaky ReLU activation to extract features. Then, two layers of convolution with kernels of 3x3, a stride of 1, and a pad of 1 are combined with a residual structure. Long and short frames are respectively processed by the feature extraction network to extract the corresponding basic features.

[0032] S6.4, design a spatial attention mechanism, splice the basic features of long and short frames extracted by the feature extraction network, pass them through two layers of convolution with 3x3 kernel, stride of 1, pad of 1, and then perform a sigmoid operation to extract the corresponding basic feature weight coefficients, and then multiply them with the corresponding basic features to form a spatial attention mechanism.

[0033] S6.5 designs a merged network using three dilated convolutional residual dense blocks concatenated together, each containing three convolutional kernels. Dilated convolutions with pads of 2 and dilation of 2 are concatenated in a densely stitched manner and then combined with residuals to form a structure. Dilated convolutions are beneficial for increasing the receptive field, and the dense stitching method is beneficial for fusing multiple convolutional features.

[0034] S6.6, design the network output, merge the network output and the basic features of the short frame, perform an addition operation, and then pass it through two convolutional kernels to... The convolution output with stride 1 and pad 1 has an output dimension of... ;

[0035] S6.7, Design the training loss, as shown in equation (4).

[0036] Equation (4)

[0037] in and These represent the width and height of the HDR image, respectively. and The curve mapping results for the HDR images generated by the model are calculated according to Equation (2), and the curve mapping results for the HDR image labels are calculated according to Equation (2). and Labels for the HDR images generated for the model and the original HDR images. Calculate the maximum value of the corresponding curve mapping result for the HDR image label according to Equation (2);

[0038] S7. Training the convolutional neural network: Train the convolutional neural network designed in step S6. The optimizer is Adam, the learning rate is 0.0001, the training period is 50, the learning rate is reduced by 0.1 times every 20 periods, and the training loss is the loss in step S6.7.

[0039] S8. Model inference post-processing and visualization:

[0040] S8.1 Post-processing of data during the inference stage: The raw data during the inference stage consists of two frames of Bayer format data, with a size of [size missing]. The data from the four channels were extracted separately to obtain... As network input, the model output is Then convert the model output back to Bayer format, and finally obtain... Synthesized HDR images;

[0041] The synthesized HDR image is visualized after curve mapping according to Equation (2) of S8.2.

[0042] Step S1, which involves acquiring raw data in a static high dynamic range scene with different suitable exposure times, and acquiring 10 frames for each scene to ensure that the synthesized HDR image can cover the dynamic range of the scene, further includes: acquiring long frame data in a static high dynamic range scene using an image sensor automatic exposure algorithm, calculating the short frame exposure time with long-to-short frame exposure ratios of 8, 12, 16, 20, 24, 28, 32, 36, 40, and acquiring short frame raw data, with a total of 10 frames acquired for each scene to ensure that the synthesized HDR image can cover the dynamic range of the scene. For example, if the image sensor automatic exposure algorithm calculates the long frame exposure time to be 1024, then 9 frames are acquired to collect short frame exposure times of 128, 85, 64, 51, 43, 37, 32, 28, 26.

[0043] Step S1 further includes:

[0044] S1.1, Based on the scene brightness, select appropriate long and short frame exposure times to ensure that bright areas in short frames are not overexposed, details in dark areas in long frames are visible, and each high dynamic range still scene can achieve the corresponding exposure. and ,in For maximum exposure time, Minimum exposure time;

[0045] S1.2, static scenes with different high dynamic ranges in and The exposure time required to obtain 10 frames of raw data is uniformly sampled. 10 frames of raw data are collected and the corresponding exposure time information is saved.

[0046] S1.3, Collect sufficient data according to steps S1.1 and S1.2.

[0047] Step S2 further includes:

[0048] S2.1, the HDR image label of each high dynamic range still scene is obtained according to the traditional weighted fusion algorithm. The traditional weighted fusion algorithm is as follows (1):

[0049] Equation (1)

[0050] in For each pixel in the raw data, The weights correspond to the pixel values. For the first Frame raw data, For the first The frame raw data corresponds to the exposure time, and gt is the HDR image label synthesized by the traditional weighted fusion algorithm; The weights of the pixel values ​​corresponding to the raw data of the i-th frame;

[0051] S2.2, For each high dynamic range still scene, perform step S2.1 to create an HDR image label for the corresponding scene.

[0052] Therefore, the advantage of this application is:

[0053] (1) The proposed HDR fusion algorithm uses only two frames, while traditional algorithms only improve the SNR Drop phenomenon after synthesizing multiple consecutive frames, and are prone to synthesizing ghosting problems. This application has relatively low requirements for image sensors and can avoid synthesizing ghosting problems as much as possible.

[0054] (2) The proposed HDR fusion algorithm does not reduce the long and short frame exposure ratio of the image sensor, does not lose the dynamic range of the scene, and maintains good details in both bright and dark areas of the scene. Attached Figure Description

[0055] The accompanying drawings, which are provided to further illustrate the invention and form part of this application, are not intended to limit the scope of the invention.

[0056] Figure 1 This is a flowchart illustrating the method described in this application. Detailed Implementation

[0057] To better understand the technical content and advantages of the present invention, the present invention will now be described in further detail with reference to the accompanying drawings.

[0058] This application provides a method for solving the SNR Drop problem after HDR image fusion, such as... Figure 1 As shown, the main implementation steps of the method are as follows:

[0059] Step S1. Data Acquisition: Acquire different exposure data under various high dynamic range static scenes. The image sensor used in this application is a Sony sensor. It can acquire two frames of raw data (one long and one short) simultaneously. In high dynamic range (HDR) static scenes, raw data is acquired with varying appropriate exposure times, 10 frames per scene, ensuring the synthesized HDR image covers the scene's dynamic range. In HDR static scenes, long frame data is acquired using the image sensor's automatic exposure algorithm. The short frame exposure time is calculated by rounding to the nearest whole number (8, 12, 16, 20, 24, 28, 32, 36, 40) and acquired as short frame raw data, 10 frames per scene, ensuring the synthesized HDR image covers the scene's dynamic range (e.g., if the image sensor's automatic exposure algorithm calculates a long frame exposure time of 1024, then 9 frames are acquired with short frame exposure times of 128, 85, 64, 51, 43, 37, 32, 28, 26).

[0060] S1.1, Based on the scene brightness, select appropriate exposure times for both long and short frames, ensuring that bright areas in short frames are not overexposed, while details in dark areas are visible in long frames, so that each high dynamic range still scene can achieve the desired effect. and ,in For maximum exposure time, Minimum exposure time;

[0061] S1.2, static scenes with different high dynamic ranges in and The exposure time required to obtain 10 frames of raw data is uniformly sampled. 10 frames of raw data are collected and the corresponding exposure time information is saved.

[0062] S1.3, Collect sufficient data according to steps S1.1 and S1.2;

[0063] Step S2. Creating HDR Image Tags

[0064] S2.1 Obtain the HDR image label for each high dynamic range still scene according to the traditional weighted fusion algorithm. The traditional weighted fusion algorithm is as follows (1):

[0065] Equation (1)

[0066] in For each pixel in the raw data, The weights correspond to the pixel values. For the first Frame raw data, For the first The frame raw data corresponds to the exposure time, and gt is the HDR image label synthesized by the traditional weighted fusion algorithm; The weights of the pixel values ​​corresponding to the raw data of the i-th frame;

[0067] S2.2 Perform step S2.1 for each high dynamic range still scene to create an HDR image label for the corresponding scene;

[0068] Step S3. Select training pair combinations

[0069] S3.1 Each high dynamic range static scene has 10 frames of raw data, which are combined in pairs. The HDR image synthesized by each combination is calculated according to Equation (1). The curve mapping result of the HDR image and the HDR image label of each combination are calculated according to Equation (2). ,

[0070] Equation (2)

[0071] in For HDR images, The constant is 5000;

[0072] S3.2 The curve mapping results corresponding to the HDR images and the curve mapping results corresponding to the HDR image labels for each combination are calculated according to formula (3). ,choose The largest combination is used as the training input data, which, together with the HDR image labels, forms a training pair combination corresponding to a high dynamic range still scene.

[0073] Equation (3)

[0074] in and These represent the width and height of the HDR image, respectively. and These are the curve mapping results for HDR images and the curve mapping results for HDR image labels, respectively.

[0075] Step S4. Data Augmentation

[0076] Raw data is generally in Bayer format. Taking RGGB Bayer as an example, the r and b channels of long and short frames are multiplied by 1.0, 1.2, 1.4, 1.6, 1.8, and 2.0 and then saved to expand the diversity of raw data. The traditional weighted fusion algorithm in step S2 is then applied to the expanded data to obtain the HDR image label of the expanded data.

[0077] Step S5. Data cropping: Crop the original raw data and HDR image labels of the training pair combination to a size of 512x512 to increase data diversity;

[0078] Step S6. Design a convolutional neural network

[0079] S6.1 Design Network Input. Since raw data is generally in Bayer format, ... Taking the rggbbayer format with a resolution as an example, the raw data is extracted sequentially from the four channels r, g, g, and b to obtain the resolution. Without losing original information, the input resolution is reduced to decrease the network computational load. The input is a superposition of two frames of data, resulting in a final network input resolution of [resolution value missing]. ;

[0080] S6.2 Design the input data preprocessing, extract the two frames of raw data into 4 channels and then normalize the data, and process the two frames of input data. and Divide by the corresponding exposure time respectively and As network input;

[0081] S6.3 Design a feature extraction network. Basic feature extraction first uses convolutional kernels... Feature extraction is performed using convolutions with a stride of 1 and a pad of 1, followed by Leaky ReLU activation. Then, two more convolutional kernels are used. A convolutional structure with stride 1 and pad 1 combined with a residual structure is used, and long and short frames are respectively processed by a feature extraction network to extract corresponding basic features.

[0082] S6.4 designs a spatial attention mechanism, splicing the basic features of long and short frames extracted by the feature extraction network, and then passing them through two convolutional kernels. After convolution with stride 1 and pad 1, a sigmoid operation is performed to extract the corresponding basic feature weight coefficients, which are then multiplied with the corresponding basic features to form a spatial attention mechanism.

[0083] The S6.5 design merges networks using three dilated convolutional residual dense blocks concatenated together, each containing three convolutional kernels. Dilated convolutions with pads of 2 and dilation of 2 are concatenated in a densely stitched manner and then combined with residuals to form a structure. Dilated convolutions are beneficial for increasing the receptive field, and the dense stitching method is beneficial for fusing multiple convolutional features.

[0084] The S6.6 network output is designed by merging the network output and the basic features of short frames, performing an addition operation, and then passing it through two convolutional kernels. The convolution output with stride 1 and pad 1 has an output dimension of... ;

[0085] S6.7 Design the training loss, as shown in equation (4).

[0086] Equation (4)

[0087] in and These represent the width and height of the HDR image, respectively. and The curve mapping results for the HDR images generated by the model are calculated according to Equation (2), and the curve mapping results for the HDR image labels are calculated according to Equation (2). and Labels for the HDR images generated for the model and the original HDR images. The maximum value of the curve mapping result corresponding to the HDR image label is calculated according to Equation (2).

[0088] Step S7. Train the convolutional neural network: Train the convolutional neural network according to the design in step S6. The optimizer is Adam, the learning rate is 0.0001, the training period is 50, the learning rate is reduced by 0.1 times every 20 periods, and the training loss is the loss in step S6.7.

[0089] Step S8. Post-processing and visualization of model inference:

[0090] S8.1 Post-processing of data during the inference stage: The raw data during the inference stage consists of two frames of Bayer format data, with a size of [size missing]. The data from the four channels were extracted separately to obtain... As network input, the model output is Then convert the model output back to Bayer format, and finally obtain... Synthesized HDR images;

[0091] The synthesized HDR image is visualized after curve mapping according to Equation (2) of S8.2.

[0092] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations can be made to the embodiments of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for solving the SNR Drop problem after HDR image fusion, characterized in that, The method includes the following steps: S1. Data Collection: Different exposure data are collected in various high dynamic range static scenes. The image sensor is used to collect two frames of raw data at a time, one long and one short. That is, raw data is collected in high dynamic range static scenes with different appropriate exposure times. 10 frames are collected for each scene to ensure that the synthesized HDR image can cover the dynamic range of the scene. S2. HDR Image Tagging Creation: HDR image labels for each high dynamic range still scene are generated based on the traditional weighted fusion algorithm. S3. Select training pair combinations: S3.1, each high dynamic range still scene has 10 frames of raw data, which are combined in pairs. The HDR image synthesized by each combination is calculated according to equation (1). The curve mapping result of the HDR image and the HDR image label of each combination are calculated according to equation (2). , Equation (2) in For HDR images, The constant is 5000; S3.2, the curve mapping results corresponding to the HDR images and the curve mapping results corresponding to the HDR image labels for each combination are calculated according to equation (3). ,choose The largest combination is used as the training input data, which, together with the HDR image labels, forms the training pair combination corresponding to the high dynamic range still scene: Equation (3) in and These represent the width and height of the HDR image, respectively. and These are the curve mapping results corresponding to HDR images and the curve mapping results corresponding to HDR image labels, respectively. S4. Data Augmentation: The raw data is in Bayer format, assuming it is rggb Bayer. The r and b channels of long and short frames are multiplied by 1.0, 1.2, 1.4, 1.6, 1.8, and 2.0 and then saved to increase the diversity of the raw data. The traditional weighted fusion algorithm in step S2 is then applied to the increased data to obtain the HDR image label of the increased data. S5. Data cropping: The original raw data and HDR image labels of the training pair combination are cropped to a size of 512x512 to increase data diversity; S6. Designing Convolutional Neural Networks: S6.1, Design the network input; the raw data is in Bayer format, assuming it is... The resolution is obtained by extracting the raw data into the rggb Bayer format, which extracts the four channels (r, g, g, b) sequentially to obtain the resolution. Without losing original information, the input resolution is reduced to decrease the network computational load. The input is a superposition of two frames of data, resulting in a final network input resolution of [resolution value missing]. ; S6.2, Design input data preprocessing: extract the two frames of raw data into four channels and then normalize the data. and Divide by the corresponding exposure time respectively and As network input; S6.3, Design the feature extraction network. Basic feature extraction first uses a 3x3 convolutional kernel with a stride of 1 and a pad of 1, followed by Leaky ReLU activation. Then, two more convolutional layers are used... A convolutional structure with stride 1 and pad 1 combined with a residual structure is used, and long and short frames are respectively processed by a feature extraction network to extract corresponding basic features. S6.4, a spatial attention mechanism is designed to concatenate the basic features of long and short frames extracted by the feature extraction network, and then pass them through two convolutional kernels. After convolution with stride 1 and pad 1, a sigmoid operation is performed to extract the corresponding basic feature weight coefficients, which are then multiplied with the corresponding basic features to form a spatial attention mechanism. S6.5 designs a merged network using three dilated convolutional residual dense blocks concatenated together, each containing three convolutional kernels. Dilated convolutions with pads of 2 and dilation of 2 are concatenated in a densely stitched manner and then combined with residuals to form a structure. Dilated convolutions are beneficial for increasing the receptive field, and the dense stitching method is beneficial for fusing multiple convolutional features. S6.6, design the network output, merge the network output and the basic features of the short frame by adding them together, and then pass it through two convolutional layers with 3x3 kernels, stride of 1, and padding of 1. The output dimension is... ; S6.7, Design the training loss, as shown in equation (4). Equation (4) in and These represent the width and height of the HDR image, respectively. and The curve mapping results for the HDR images generated by the model are calculated according to Equation (2), and the curve mapping results for the HDR image labels are calculated according to Equation (2). and Labels for the HDR images generated for the model and the original HDR images. Calculate the maximum value of the corresponding curve mapping result for the HDR image label according to Equation (2); S7. Training the convolutional neural network: Train the convolutional neural network designed in step S6. The optimizer is Adam, the learning rate is 0.0001, the training period is 50, the learning rate is reduced by 0.1 times every 20 periods, and the training loss is the loss in step S6.

7. S8. Model inference post-processing and visualization: S8.1 Post-processing of data during the inference stage: The raw data during the inference stage consists of two frames of Bayer format data, with a size of [size missing]. The data from the four channels were extracted separately to obtain... As network input, the model output is Then convert the model output back to Bayer format, and finally obtain... Synthesized HDR images; The synthesized HDR image is visualized after curve mapping according to Equation (2) of S8.

2.

2. The method for solving the SNR Drop problem after HDR image fusion according to claim 1, characterized in that, Step S1, which involves acquiring raw data in a static high dynamic range scene with different suitable exposure times, collecting 10 frames for each scene to ensure that the synthesized HDR image can cover the dynamic range of the scene, further includes: acquiring long frame data in a static high dynamic range scene using an image sensor automatic exposure algorithm, calculating the short frame exposure time with long-to-short frame exposure ratios of 8, 12, 16, 20, 24, 28, 32, 36, 40 (rounded), and acquiring short frame raw data, collecting a total of 10 frames for each scene to ensure that the synthesized HDR image can cover the dynamic range of the scene. For example, assuming the image sensor automatic exposure algorithm calculates the long frame exposure time to be 1024, then acquiring 9 more frames with short frame exposure times of 128, 85, 64, 51, 43, 37, 32, 28, 26.

3. The method for solving the SNR Drop problem after HDR image fusion according to claim 1, characterized in that, Step S1 further includes: S1.1, Based on the scene brightness, select appropriate long and short frame exposure times to ensure that bright areas in short frames are not overexposed, details in dark areas in long frames are visible, and each high dynamic range still scene can achieve the corresponding exposure. and , For maximum exposure time, Minimum exposure time; S1.2, static scenes with different high dynamic ranges in and The exposure time required to obtain 10 frames of raw data is uniformly sampled. 10 frames of raw data are collected and the corresponding exposure time information is saved. S1.3, Collect sufficient data according to steps S1.1 and S1.

2.

4. The method for solving the SNR Drop problem after HDR image fusion according to claim 1, characterized in that, Step S2 further includes: S2.1, the HDR image label of each high dynamic range still scene is obtained according to the traditional weighted fusion algorithm. The traditional weighted fusion algorithm is as follows (1): Equation (1) in For each pixel in the raw data, The weights correspond to the pixel values. For the first Frame raw data, For the first The frame raw data corresponds to the exposure time, and gt is the HDR image label synthesized by the traditional weighted fusion algorithm; The weights of the pixel values ​​corresponding to the raw data of the i-th frame; S2.2, For each high dynamic range still scene, perform step S2.1 to create an HDR image label for the corresponding scene.

Images