Low-light image enhancement method suitable for haze climate

Abstract

The invention relates to a low-light image enhancement method suitable for haze climate, and belongs to the field of digital image processing. According to the method, firstly, denoising information in an image pair can be effectively learned in the learning process through a GCA (Group Channel Attention) group channel attention structure, and interference caused by external noise can be solved; secondly, the transformer block of the method enables the network to have a better window adaptive capability through a self-attention mechanism, and has a better effect in solving a local overexposure phenomenon; moreover, according to the method, fine adjustment can be carried out through a small number of field samples, the adaptive capacity to the required environment is greatly improved, a new task can be learned on a low-light enhancement task, and the influence of non-illumination factors on the image quality is solved.

Description

Technical Field

[0001] The invention belongs to the field of digital image processing, and in particular relates to a low-light image enhancement method suitable for haze climate. Background Art

[0002] Low-light enhancement is a technology that acquires, converts, enhances, records and displays target image information under low-light conditions. However, the environment itself is affected by other non-low-light factors, such as fog on the sea at night, smog in the city at night, and rainy environment. Therefore, when performing low-light enhancement in a haze environment, it is necessary to consider both low-light enhancement and defogging.

[0003] At present, the field of low-light enhancement has made many major breakthroughs in the field of deep learning for problems such as external light exposure, noise, color cast, etc. through different reliance on data sets and learning methods. However, from the perspective of existing practical scenarios, there are still many factors that have not been considered, and the same is true for dehazing.

[0004] As key sub-fields of image restoration, low-light enhancement and dehazing play a wide range of roles in the field of computer vision. They can better capture the problem of missing image information caused by dim image light and fog, and perform restoration based on corresponding methods.

[0005] Commonly used traditional methods for low-light enhancement include: image equalization, gamma correction, Retinex algorithm, etc., and commonly used traditional methods for defogging include: dark channel defogging, etc. In addition, many deep learning methods are gradually applied to low-light enhancement and defogging tasks. Recent studies have shown that deep learning-based methods can significantly improve the visual quality of images under low-light or foggy conditions. For example, some methods can effectively reduce noise and enhance details, making the image visually clearer and more natural. With the further development of deep learning technology and the improvement of computing hardware, low-light image enhancement is expected to be used in a wider range of application scenarios, including night photography, security monitoring, unmanned driving and other fields. At the same time, continuing to improve the efficiency and performance of the algorithm, as well as further exploring new deep learning architectures and training methods, will be the focus of future research.

[0006] There are three main problems in the existing deep learning methods for low-light image enhancement:

[0007] 1) Real low-light images are more susceptible to external additive noise due to their darker colors, which leads to the additive noise being considered as useful information during image enhancement.

[0008] 2) Real low-light images are often accompanied by the appearance of light sources. Due to the limited tolerance of camera sensors, it is impossible to record both the dark environment and the luminous light source at the same time, resulting in local overexposure problems in night scene photography.

[0009] 3) The environment itself is affected by other non-twilight factors, such as fog on the sea at night, smog in the city at night, and rainy environment. Affected by these non-twilight factors, using only a general low-light enhancement network often cannot constrain this influence. Training two networks separately to complete low-light image enhancement and defogging tasks and then splicing them will lead to the problem of doubling model parameters. Summary of the invention

[0010] The purpose of the present invention is to overcome the problems existing in the above-mentioned prior art and provide a low-light image enhancement method suitable for haze climate. Firstly, the method can effectively learn the denoising information in the image pair during the learning process through the GCA (Group Channel Attention) group channel attention structure, and can solve the interference caused by external noise; secondly, the transformer block of the method enables the network to have better window adaptation ability through the self-attention mechanism, and can have better effect in solving the local overexposure phenomenon; furthermore, the method can be fine-tuned through a small number of on-site samples, which greatly improves the adaptability to the required environment, and can learn new tasks on the low-light enhancement task to solve the influence of non-light factors on image quality.

[0011] To achieve the above-mentioned purpose, the technical solution of the present invention is: a low-light image enhancement method suitable for haze climate, which adopts self-attention mechanism to self-allocate image weights, and introduces Unet structure to obtain multi-scale features to realize adaptive adjustment of windows during low-light image enhancement and defogging tasks, and adopts continuous learning training method to coordinate low-light image enhancement and defogging tasks.

[0012] In one embodiment of the present invention, a low-light image enhancement and defogging network is constructed, which consists of two parts: the first part is an improved Transformer structure and a self-encoder / decoder composed of the improved Transformer structure, and the second part is a plug-and-play module composed of TUnet+GCA; finally, a continuous learning training method is adopted to make the low-light image enhancement and defogging network suitable for foggy climate.

[0013] In one embodiment of the present invention, the low-light image enhancement and defogging network structure is expressed as:

[0014]

[0015] Among them, a1, a2, b1, b2, c1, c2, and res are all intermediate variables, T1 and T2 represent single-layer TransformerBlock and double-layer Transformer Block respectively, F represents the feature fusion module, GCA is the group channel attention structure, Conv represents the convolution operation, the subscripts 1*1 and 3*3 represent 1*1 convolution kernel and 3*3 convolution respectively, and input represents input.

[0016] In one embodiment of the present invention, the improved Transformer structure and the self-codec composed of the improved Transformer structure are specifically as follows:

[0017] Improved Transformer structure, using a ternary gated feedforward network;

[0018] The improved Transformer structure is used to construct the front-end and back-end self-encoder-decoders, that is, three Transformer blocks, a feature fusion module, and a convolutional layer are used to construct the self-encoder-decoder.

[0019] In one embodiment of the present invention, the feature fusion module is composed of three convolutional layers.

[0020] In one embodiment of the present invention, a plug-and-play module composed of TUnet+GCA is used, as follows:

[0021] The improved Transformer structure replaces the original convolutional layer in Unet to form a TUnet structure, that is, the four downsamplings and one upsampling of the TUnet front-end structure and the jump connections at the input and output ends are embedded with the improved Transformer structure of the predetermined number of layers, and the output result of the TUnet front-end structure is connected to the TUnet back-end structure through the jump connection with parameters. The last matrix summation operation sums the three parameterized branches and sends them to the GCA structure;

[0022] The GCA structure contains multiple skip links and concat operations.

[0023] In one embodiment of the present invention, the training method of continuous learning is as follows:

[0024] Step 1. Adjust the low-light image enhancement and defogging network model structure according to the GPU size to ensure that CUDA out of memory does not occur.

[0025] Step 2: Perform data preprocessing operations including random cropping and shuffling on the data set required for training;

[0026] Step 3: Introduce a paired dataset of low-light enhancement into the low-light image enhancement and dehazing network model, train the main task, i.e., the low-light image enhancement task, and obtain the pre-training weights, which are named Params1 here;

[0027] Step 4. Import nighttime pictures or high-definition pictures of the required application scenarios;

[0028] Step 5. If there is no HD image, perform post-stage degradation and pre-stage enhancement on the night image. If there is no night image and the HD image has reached the optimal state and does not need to be enhanced, only perform post-stage degradation on the HD image to obtain a new data set, which is named Data1.

[0029] Step 6, training the sub-task, i.e. the defogging task. The data set uses the defogging paired data set and the data set processed by the previous and next stages in Step 4. A regularization constraint term is introduced into the original loss in the sub-task training. The Euclidean distance between the pre-trained weights Params2 and Params1 in the sub-task training is constrained by regularization.

[0030] Step 7. If you need to continue training new image restoration tasks, repeat Step 4-Step 6; otherwise, execute Step 8;

[0031] Step 8: The final output result is obtained through post-processing including lighting correction and white balance.

[0032] In one embodiment of the present invention, the post-stage degradation model uses a fogging algorithm, dark processing and APSF blur kernel for light blurring; the pre-stage enhancement model uses the main task model for enhancement, and the obtained result is subjected to filtering and denoising.

[0033] In one embodiment of the present invention, the overall loss function of the method is as follows:

[0034] Loss=α×L1+β×L2+γ×(1-s)+L3

[0035]

[0036] Among them, α, β, γ, and weight are hyperparameters, result and label are predicted values ​​and true values, SSIM is structural similarity, and params1 and params2 are model weights of the main task and sub-task;

[0037] The loss term L1 is the edge loss term of the Sobel operator. It is obtained by using two types of Sobel operators, the horizontal axis and the vertical axis, to extract the edge information of the output result and the high-definition image, and using both as the input of the SmoothL1 Loss (smooth L1 loss function);

[0038] The loss term L2 directly uses the network output and high-definition image as the input of SmoothL1 Loss;

[0039] The loss term 1-s is the SSIM loss, which is obtained by subtracting the SSIM index of the output result and the high-definition image from 1;

[0040] The loss term L3 is the continuous learning regularization loss, which is used to constrain the distance between the parameters learned when learning the sub-task and the parameters learned in the old task, thereby reducing catastrophic forgetting.

[0041] The present invention also provides a computer-readable storage medium, on which computer program instructions that can be executed by a processor are stored. When the processor executes the computer program instructions, the method steps described above can be implemented.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] ① Compared with the existing deep learning methods, the present invention does not increase the computational complexity of the model testing and reasoning stages during the collaborative low-light enhancement and defogging tasks, thereby reducing the training and reasoning burden of the GPU.

[0044] ② Compared with the existing end-to-end deep learning method, the image quality of the output results of the present invention exceeds that of the GT images of the single task of low-light enhancement and the single task of defogging, breaking the method of stacking image quality only by high-quality data sets.

[0045] ③Compared with existing traditional methods, the innovation of GCA structure and Transformer's ternary gated feedforward network enables the learning process to effectively utilize intermediate features, which is effective in solving external noise and local exposure problems.

[0046] ④ Compared with existing training methods, the present invention can fine-tune the model through a small number of application scenario samples to adapt to different scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is the overall network framework of the present invention.

[0048] Figure 2 Schematic diagram of the improved Transformer structure and the self-encoder-decoder structure composed of it.

[0049] Figure 3 Schematic diagram of the GCA structure.

[0050] Figure 4 The overall flow chart of the continuous learning training strategy.

[0051] Figure 5 This is a schematic diagram of the front stage enhancement and the back stage degradation. DETAILED DESCRIPTION

[0052] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings.

[0053] The present invention provides a low-light image enhancement method suitable for haze climate, adopts self-attention mechanism to self-allocate image weights, introduces Unet structure to obtain multi-scale features to realize adaptive adjustment of windows during low-light image enhancement and defogging tasks, and adopts continuous learning training method to coordinate low-light image enhancement and defogging tasks.

[0054] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0055] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0056] like Figure 1 As shown, this embodiment provides a low-light image enhancement method suitable for haze climate. The overall network framework is as follows Figure 1 As shown in the figure, the self-attention mechanism is used to self-allocate image weights, and the Unet structure is introduced to obtain multi-scale features to achieve adaptive adjustment of windows during low-light image enhancement and defogging tasks, and a continuous learning training method is used to coordinate low-light image enhancement and defogging tasks. The entire network consists of two parts: the first part is the network's self-encoder and decoder and the improved Transformer structure therein, and the second part is a plug-and-play module composed of TUnet+GCA. Finally, a continuous learning training method is used to make the low-light enhancement method applicable to foggy climates.

[0057] In order to allow the network to acquire more relevant knowledge of the previous and next stages during the learning process and speed up the network convergence, a large number of RL (Resnet Learning) operations are introduced in the previous and next stages of the network structure, so that the network can capture the commonalities between different tasks more quickly during continuous learning and training. The network structure of this application is represented as:

[0058]

[0059] Among them, a1, a2, b1, b2, c1, c2, and res are all intermediate variables, T1 and T2 represent single-layer TransformerBlock and double-layer Transformer Block respectively, F represents the feature fusion module, GCA is the group channel attention structure, Conv represents the convolution operation, the subscripts 1*1 and 3*3 represent 1*1 convolution kernel and 3*3 convolution respectively, and input represents input.

[0060] Detailed Design

[0061] 1. Model structure

[0062] 1) Improved Transformer structure and self-encoder / decoder composed of it (such as Figure 2 (shown)

[0063] The network model adopts a new Transformer structure. Compared with the binary gated feedforward network used in the past, the ternary gated feedforward network is easy to learn more intermediate features. The function that could only learn quadratic correlation is raised to a maximum exponential of three times, and three adaptive parameters are introduced to realize the ternary gated feedforward network.

[0064] The present invention uses the Transformer structure to construct the self-encoder and decoder of the front and back stages, and respectively uses three Transformer blocks, a feature fusion module, and a convolutional layer to construct the self-encoder and decoder of the present invention. In addition, the feature fusion module of the self-encoder and decoder is composed of three convolutional layers.

[0065] 2)TUnet+GCA

[0066] The TUnet structure of the present invention replaces the original convolutional layer in Unet with the new Transformer structure designed by the present invention, embeds a certain number of Transformer structures in the four downsamplings and one upsampling of the front-stage structure of TUnet, and the jump connections at the input and output ends, and connects the output result of the front stage with the back stage through the jump connection with parameters. The last matrix summation operation sums the three parameterized branches and sends them to the GCA structure.

[0067] The GCA structure (such as Figure 3 As shown in the figure, it contains multiple skip links and concat operations, which is conducive to network learning to retain rich low-frequency information in the image. Since external additive noise usually exists in the image at a higher frequency, the GCA structure can effectively learn the denoising information in the image pair during the learning process, and can better solve the problem 1) The interference caused by external noise.

[0068] 2. Continuous learning and training strategy

[0069] Step 1: Adjust the network model structure according to the size of your own GPU to ensure that CUDA out of memory does not occur.

[0070] Step 2: Perform data preprocessing operations such as random cropping and shuffling on the data set required for training.

[0071] Step 3: Introduce paired datasets of low-light enhancement (LOL, MIT, etc.) into the network model, train the main task, and obtain pre-trained weights, named Params1 here.

[0072] Step 4: Introduce night pictures or high-definition pictures of the required application scenarios. There is no need for paired data sets. The data set we use is night pictures.

[0073] Step 5: If there is no HD image, perform post-stage degradation and pre-stage enhancement on the night image. If there is no night image and the HD image has reached the optimal state and does not need to be enhanced, only perform post-stage degradation on the HD image. Thus, a new data set is obtained, which is named Data1. (The post-stage degradation model uses the traditional fogging algorithm, dark processing and APSF blur kernel for light blurring; the pre-stage enhancement model uses the main task model for enhancement, and appropriately uses filtering and denoising processing on the obtained results [depending on the shooting situation, if the noise has no effect, the denoising part can be ignored to avoid losing useful high-frequency information])

[0074] Step 6: Training sub-task, i.e., dehazing task. The data set uses the dehazing paired data set (Dehaze_out, NH-HAZE, etc.) and the data set after the front-end and back-end processing in Step 4. A regularization constraint term is introduced into the original loss of this training. The Euclidean distance between the pre-trained weights Params2 and Params1 of this training is constrained by regularization.

[0075] Step 7: If you need to continue training new image restoration tasks (such as deblurring, rain removal, super-resolution, etc.), repeat Step 4-6.

[0076] Step 8: After post-processing such as illumination correction and white balance, the final output result is obtained. (For the overall flow chart, refer to Figure 4 , reference for front stage enhancement and back stage degradation Figure 5 ).

[0077] Since the network of the present invention can be fine-tuned with a small number of field samples, the adaptability to the required environment is greatly improved, and new tasks can be learned on the low-light enhancement task. In solving problem 3) the influence of non-lighting factors on image quality has a great advantage and the effect is significant.

[0078] 3. Loss Function

[0079] This training method is to ensure that the image can not only reduce the error at the pixel level during the restoration process, but also have a good restoration effect in terms of structure and illumination, so that our experimental results have high PSNR and SSIM indicators while also having good visual effects. Therefore, the following expression is used as the loss function of our experiment, where α, β, γ, and weight are hyperparameters, result and label are predicted values ​​and true values, SSIM is structural similarity, and params1 and params2 are the model weights of the main task and sub-task:

[0080] Loss=α×L1+β×L2+γ×(1-s)+L3

[0081] The loss function consists of four items, which can be expressed as:

[0082]

[0083] The loss term L1 is the edge loss term of the Sobel operator. It is obtained by using two types of Sobel operators, the horizontal axis and the vertical axis, to extract the edge information of the output result and the high-definition image, and using both as the input of the SmoothL1 Loss (smooth L1 loss function). This loss introduces implicit edge prior knowledge, so that the network can better learn the edge information of the image during the training process.

[0084] The loss term L2 directly uses the network output result and the high-definition image as the input of SmoothL1 Loss, which can better obtain the loss of the image at the pixel level.

[0085] The loss term 1-s is the SSIM loss, which is obtained by subtracting the SSIM index of the output result and the high-definition image from 1. This loss can make the network result have better structural similarity.

[0086] The loss term L3 is the continuous learning regularization loss, which is used to constrain the distance between the parameters learned by the new model when learning the sub-task and the parameters learned by the old task, thereby reducing catastrophic forgetting.

[0087] The present invention also provides a computer-readable storage medium, on which computer program instructions that can be executed by a processor are stored. When the processor executes the computer program instructions, the method steps described above can be implemented.

[0088] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0089] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0090] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0091] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0092] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. A low-light image enhancement method suitable for haze climate, characterized in that: The self-attention mechanism is used to self-assign image weights, and the Unet structure is introduced to obtain multi-scale features to achieve adaptive adjustment of windows during low-light image enhancement and defogging tasks. At the same time, a continuous learning training method is adopted to coordinate low-light image enhancement and defogging tasks.

2. The low-light-level image enhancement method applicable to haze climate according to claim 1, characterized in that: A low-light image enhancement and dehazing network is constructed, which consists of two parts: the first part is an improved Transformer structure and a self-encoder-decoder composed of the improved Transformer structure; the second part is a plug-and-play module composed of TUnet+GCA; finally, a continuous learning training method is used to make the low-light image enhancement and dehazing network suitable for foggy climates.

3. The low-light-level image enhancement method applicable to haze climate according to claim 2, characterized in that: The low-light image enhancement and dehazing network structure is expressed as: Among them, a1, a2, b1, b2, c1, c2, and res are all intermediate variables, T1 and T2 represent single-layer Transformer Block and double-layer Transformer Block, respectively, F represents the feature fusion module, GCA is the group channel attention structure, Conv represents the convolution operation, the subscripts 1*1 and 3*3 represent 1*1 convolution kernel and 3*3 convolution, respectively, and input represents the input.

4. The low-light-level image enhancement method applicable to haze climate according to claim 2, characterized in that: The improved Transformer structure and the self-encoder-decoder composed of the improved Transformer structure are as follows: Improved Transformer structure, using a ternary gated feedforward network; The improved Transformer structure is used to construct the front-end and back-end self-encoder-decoders, that is, three Transformer blocks, a feature fusion module, and a convolutional layer are used to construct the self-encoder-decoder.

5. The low-light-level image enhancement method applicable to haze climate according to claim 4, characterized in that: The feature fusion module consists of three convolutional layers.

6. The low-light-level image enhancement method applicable to haze climate according to claim 4, characterized in that: The plug-and-play module composed of TUnet+GCA is as follows: The improved Transformer structure replaces the original convolutional layer in Unet to form the TUnet structure. That is, the improved Transformer structure with a predetermined number of layers is embedded in the four downsampling and one upsampling of the TUnet front-end structure and the jump connection at the input and output ends. The output result of the TUnet front-end structure is connected to the TUnet back-end structure through the jump connection with parameters. The last matrix summation operation sums the three parameterized branches and sends them to the GCA structure. The GCA structure contains multiple skip links and concat operations.

7. The low-light-level image enhancement method applicable to haze climate according to claim 1 or 2, characterized in that: The training methods for continuous learning are as follows: Step 1: Adjust the low-light image enhancement and dehazing network model structure according to the GPU size to ensure that there is no shortage of video memory; Step 2: Perform data preprocessing operations on the training data set, including random cropping and shuffling; Step 3: Introduce a paired dataset of low-light enhancement into the low-light image enhancement and dehazing network models, train the main task, i.e., the low-light image enhancement task, and obtain pre-trained weights, here named Params1; Step 4. Import nighttime pictures or high-definition pictures of the required application scenario; Step 5: If there is no HD image, perform post-stage degradation and pre-stage enhancement on the night image. If there is no night image and the HD image has reached the optimal state and does not need enhancement, only perform post-stage degradation on the HD image to obtain a new dataset, which is named Data1. Step 6: Train the sub-task, i.e., the defogging task. The dataset uses the defogging paired dataset and the dataset processed by the previous and next stages in Step 4. A regularization constraint is introduced into the original loss in the sub-task training. The Euclidean distance between the pre-trained weights Params2 and Params1 in the sub-task training is constrained by regularization. Step 7. If you need to continue training new image restoration tasks, repeat Step 4-Step 6. Otherwise, execute Step 8; Step 8: The final output result is obtained through post-processing including lighting correction and white balance.

8. The low-light-level image enhancement method applicable to haze climate according to claim 7, characterized in that: The post-stage degradation model uses atomization algorithm, dark processing and APSF blur kernel for light blurring; the pre-stage enhancement model uses the main task model for enhancement, and the obtained result is subjected to filtering and denoising.

9. The low-light-level image enhancement method applicable to haze climate according to claim 1, characterized in that: The overall loss function of this method is as follows: Loss=α×L1+β×L2+γ×(1-s)+L3 Among them, α, β, γ, and weight are hyperparameters, result and label are predicted values ​​and true values, SSIM is structural similarity, params1 and params2 are the model weights of the main task and sub-task; The loss term L1 is the edge loss term of the Sobel operator. It is obtained by using two types of Sobel operators, the horizontal axis and the vertical axis, to extract the edge information of the output result and the high-definition image, and using both as the input of the smooth L1 loss function SmoothL1 Loss. The loss term L2 directly uses the network output and high-definition image as the input of SmoothL1 Loss; The loss term 1-s is the SSIM loss, which is obtained by subtracting the SSIM index of the output result and the high-definition image from 1; The loss term L3 is the continuous learning regularization loss, which is used to constrain the distance between the parameters learned when learning the sub-task and the parameters learned in the old task, thereby reducing catastrophic forgetting.

10. A computer-readable storage medium storing computer program instructions that can be executed by a processor, wherein when the processor executes the computer program instructions, the method steps according to any one of claims 1 to 9 can be implemented.

Images