Image processing methods, devices and storage media
By employing a dual-path parallel processing architecture and feature consistency loss optimization, the problem of feature distortion in diffusion model cleanup methods is solved, achieving stable cleanup of adversarial images and full utilization of classifier knowledge, thus ensuring the accuracy and consistency of image processing.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HARBIN INST OF TECH
- Filing Date
- 2026-01-16
- Publication Date
- 2026-06-02
AI Technical Summary
Existing diffusion-based cleaning methods may introduce new feature distortions during the denoising process, leading to inconsistent feature representations in image classifiers and insufficient utilization of the discriminative knowledge of the target classifier, resulting in instability.
A dual-path parallel processing architecture is adopted. By using a pre-trained target classifier and diffusion model, the direct path feature representation and the diffusion path feature representation are obtained. The image update strategy is optimized by using feature consistency loss. Combined with ADAM optimizer and range cropping, adversarial image purification is achieved.
By maintaining the stability of the purification process through multiple iterations and making full use of classifier knowledge, effective purification of adversarial images can be achieved, ensuring the accuracy and consistency of image processing.
Smart Images
Figure CN122135064A_ABST
Abstract
Description
Technical Field
[0001] This application relates to an image processing method, an electronic device, and a computer-readable storage medium, belonging to the field of information attack and defense technology. Background Technology
[0002] Currently, attacks on neural networks used for vision tasks can be achieved by adding tiny perturbations to the input image that are almost imperceptible to the naked eye, causing the neural network to misclassify.
[0003] Among current adversarial defense techniques against the aforementioned attacks, descaling methods are widely used because they do not require model retraining. For example, descaling methods based on diffusion models can generate high-quality images through a stepwise denoising process and utilize prior knowledge to remove adversarial perturbations imposed by the attacker. However, current descaling methods based on diffusion models may introduce new feature distortions during the denoising process, leading to inconsistent feature representations of the image in the classifier and insufficient utilization of the discriminative knowledge of the target classifier, resulting in insufficient stability of the descaling methods. Summary of the Invention
[0004] This application discloses an image processing method, an electronic device, and a computer-readable storage medium.
[0005] The image processing method in this application is based on a pre-trained target classifier and a pre-trained diffusion model. The method specifically includes: Based on the target classifier and the diffusion model, direct path feature representation and diffusion path feature representation are obtained according to the initial adversarial image or the current adversarial image at the end of the previous iteration. The current adversarial image is determined by updating the initial adversarial image through one or more iterations. The direct path feature representation and the diffusion path feature representation have the same number of feature levels, which are determined by the target classifier. Based on the feature hierarchy, the feature consistency loss is determined according to the direct path feature representation and the diffusion path feature representation; Using the feature consistency loss as an optimization benchmark, an image update strategy is determined to update the current adversarial image at the end of the current iteration.
[0006] In some implementations, obtaining direct path feature representations and diffusion path feature representations based on the target classifier and the diffusion model, according to the initial adversarial image or the current adversarial image at the end of the previous iteration, includes: Based on the target classifier, the direct path feature representation is determined according to the initial adversarial image or the current adversarial image at the end of the previous iteration; Based on the initial adversarial image or the current adversarial image at the end of the previous iteration, perform Gaussian noise simulation diffusion to determine the current noise image; Based on the diffusion model, denoising and reconstruction are performed according to the current noisy image to determine the reconstructed image; Based on the target classifier, the diffusion path feature representation is determined according to the reconstructed image.
[0007] In some implementations, the step of performing Gaussian noise simulation diffusion based on the initial adversarial image or the current adversarial image at the end of the previous iteration to determine the current noise image is as follows:
[0008] in For the current noisy image, The initial adversarial image or the current adversarial image at the end of the previous iteration. The amount of Gaussian noise to be added. The attenuation coefficient is... For noise intensity, It is an identity matrix.
[0009] In some implementations, determining the feature consistency loss based on the feature hierarchy, according to the direct path feature representation and the diffusion path feature representation, includes: Based on the feature hierarchy, the direct path feature representation and the diffusion path feature representation are aligned so that the direct path feature representation and the diffusion path feature representation with the same feature hierarchy correspond to each other. The initial feature consistency loss is determined based on the corresponding direct path feature representations and diffusion path feature representations. Based on the weight coefficients corresponding to different feature levels, the initial feature consistency loss function is modified to determine the feature consistency loss.
[0010] In some implementations, the initial feature consistency loss is determined based on the corresponding direct path feature representations and diffusion path feature representations, using the following formula:
[0011] in The initial feature consistency loss, L The total number of the feature levels. This refers to the direct path feature representation corresponding to the feature level. This represents the diffusion path feature corresponding to the feature level. For the reconstructed image; The initial feature consistency loss function is modified according to the weight coefficients corresponding to different feature levels to determine the feature consistency loss, as shown in the formula:
[0012] in The feature consistency loss, The weight coefficients are those corresponding to the feature level.
[0013] In some implementations, determining an image update strategy based on the feature consistency loss to update the current adversarial image at the end of the current iteration includes: Based on the ADAM optimizer, an initial image update strategy is determined according to the feature consistency loss, the initial adversarial image or the current adversarial image at the end of the previous iteration, and a preset learning rate, so as to update the original current adversarial image at the end of the current iteration. Based on the preset adversarial perturbation amplitude and the initial adversarial image, the original current adversarial image is cropped to determine the current adversarial image at the end of the current iteration.
[0014] In some implementations, the ADAM-based optimizer determines the initial image update strategy based on the feature consistency loss, the initial adversarial image or the current adversarial image at the end of the previous iteration, and a preset learning rate, using the following formula:
[0015] in This refers to the ADAM optimizer. The current adversarial image at the end of the last iteration. This is the original current adversarial image at the end of the current iteration. For the initial adversarial image, The learning rate of the ADAM optimizer. For gradient operators, The feature consistency loss is given.
[0016] In some implementations, the process of cropping the original current adversarial image based on a preset adversarial perturbation amplitude and the initial adversarial image to determine the current adversarial image at the end of the current iteration is described by the following formula:
[0017] in, This represents the current adversarial image at the end of the current iteration, where `clip` is the truncation function. This is the preset counter-disturbance amplitude.
[0018] The electronic device in this application includes a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the image processing method described above is implemented.
[0019] The computer-readable storage medium in this application embodiment stores a computer program that, when executed by the processor, implements the image processing method described above.
[0020] The beneficial effects of this application are as follows: By establishing a dual-path parallel processing architecture, this application utilizes the aforementioned dual-path parallel processing in multiple iterations of the image processing process. In the classifier, the consistency difference between natural images and adversarial image samples in the feature space is determined based on the input adversarial image samples. Furthermore, multi-level feature consistency constraints are used to optimize the feature representation. Finally, iterative optimization is performed based on gradient descent to process the input adversarial image samples, thereby achieving purification processing of adversarial image samples. While achieving the processing objective, the stability of the purification process and the full utilization of classifier knowledge are maintained. Attached Figure Description
[0021] Figure 1 This is one of the flowcharts illustrating the image processing method in the embodiments of this application; Figure 2 This is a second schematic flowchart of the image processing method in the embodiments of this application; Figure 3 This is the third flowchart illustrating the image processing method in the embodiments of this application; Figure 4 This is the fourth flowchart illustrating the image processing method in the embodiments of this application. Detailed Implementation
[0022] Please see Figure 1 The image processing method in this application is based on a pre-trained target classifier and a pre-trained diffusion model. The method specifically includes the following steps: Step 01: Based on the target classifier and the diffusion model, obtain the direct path feature representation and the diffusion path feature representation according to the initial adversarial image or the current adversarial image at the end of the previous iteration. The current adversarial image is determined by updating the initial adversarial image through one or more iterations. The direct path feature representation and the diffusion path feature representation have the same number of feature levels, and the feature levels are determined by the target classifier. Step 02: Based on the feature hierarchy, determine the feature consistency loss according to the direct path feature representation and the diffusion path feature representation; Step 03: Using feature consistency loss as the optimization benchmark, determine the image update strategy to update the current adversarial image at the end of the current iteration.
[0023] Specifically, the image processing method in this application is used for adversarial defense against attacks on neural networks used for visual tasks. The main approach involves purifying the initial adversarial image before it is input into the neural network to eliminate subtle perturbations imperceptible to the human eye, thereby protecting the neural network. This image processing method is based on a pre-trained target classifier and a diffusion model. For example, the target classifier can be a model such as CIFAR-10 or ResNet-50, and the diffusion model can be a model such as DDPM or DDIM. In the image processing method described above, only the progressive denoising function of the diffusion model is used.
[0024] Before executing the image processing method described above, initialization parameter configuration is first performed for the target classifier and the diffusion model, and relevant parameters for the overall iterative process of the image processing method are also initialized. For example, for the target classifier, the hierarchical set of extractable features of the target classifier is initialized. L And the weight coefficients of each feature layer, for example, for the CIFAR-10 model, which includes a set of feature layers. L It includes three layers: shallow (block1), middle (block2), and deep (block3), with corresponding weight coefficients of 0.2, 0.3, and 0.5, respectively. The ResNet-50 model includes a set of feature layers... L It includes four layers: shallow (conv1), mid-shallow (layer1), middle (layer2), and deep (layer3), with corresponding weight coefficients of 0.1, 0.2, 0.3, and 0.4, respectively. Furthermore, for the process of performing range clipping using the target classifier, the adversarial perturbation amplitude is initialized. For example, for the CIFAR-10 model, the adversarial perturbation amplitude is typically initialized to... .
[0025] Furthermore, for the process of adding Gaussian noise, for example, the noise intensity is initialized to 0.05~0.2, the attenuation coefficient is initialized to 0.8~1.0, and the learning rate of the ADAM optimizer is initialized to 0.01~0.2. For the overall iterative process, the maximum number of iterations is initialized to 5~20, and the consistency loss threshold is initialized to 0.05. After the above parameters are initialized, the image processing method described in the above embodiment can be executed. It should be noted that the following examples illustrate the image processing method in the above embodiment using the steps executed in a single iteration; the steps executed in each iteration are the same.
[0026] In some implementations, before the start of the current iteration, if it is the first iteration, an initial adversarial image can be obtained; if it is not the first iteration, the current adversarial image obtained at the end of the previous iteration can be obtained. Then, based on the actual situation of the current iteration, one of these two images is used as the input image for image processing. First, based on the target classifier and the diffusion model, a dual-path feature extraction method is used to extract the direct path feature representation and the diffusion path feature representation corresponding to the input image, respectively. The goal of dual-path feature extraction is to construct two semantically equivalent but different processing paths for each input image, fully utilizing the discriminative knowledge within the target classifier to facilitate further use of feature consistency constraints to confirm the optimization target of image processing.
[0027] In some implementations, please refer to Figure 2 Step 01 includes: Step 011: Based on the target classifier, determine the direct path feature representation according to the initial adversarial image or the current adversarial image at the end of the previous iteration; Step 012: Based on the initial adversarial image or the current adversarial image at the end of the previous iteration, perform Gaussian noise simulation diffusion to determine the current noise image; Step 013: Based on the diffusion model, perform denoising and reconstruction according to the current noisy image to determine the reconstructed image; Step 014: Based on the target classifier, determine the feature representation of the diffusion path according to the reconstructed image.
[0028] It should be noted that the execution order of step 011 is not limited to... Figure 2 As shown, steps 011 and 012-014 can be executed in parallel or in any order. Figure 2 This is for illustrative purposes only and should not be interpreted as a limitation on the order of execution.
[0029] Specifically, please refer to the following examples for the dual-path feature extraction process described above. In some examples, the extraction of direct path feature representations is generally achieved by calling the target classifier, directly using the input image of the current iteration as the data basis, and adjusting the extraction based on the selected level of the target classifier. l Perform the corresponding extraction. For example, taking the CIFAR-10 model as an example, then... l We can choose blocks 1, 2, and 3. The shallow layer (block 1) corresponds to the image edges, the middle layer (block 2) corresponds to the image texture, and the deep layer (block 3) corresponds to the image semantic information. Based on the CIFAR-10 model, assuming the previous iteration corresponds to the number of iterations... (like (meaning the current iteration is the first iteration), then the direct path feature representation obtained directly through the CIFAR-10 model can be represented as: ,in Representing the CIFAR-10 model, This represents the current adversarial image at the end of the last iteration. Specifically, when... hour This represents the initial adversarial image.
[0030] To represent the diffusion path features, Gaussian noise needs to be added to the input image to simulate the diffusion process. The specific noise addition process can be shown in Equation 1.
[0031] …………Formula 1 in For the current noisy image, The initial adversarial image or the current adversarial image at the end of the last iteration. The attenuation coefficient is... For noise intensity, It is an identity matrix. The amount of Gaussian noise to be added. It follows a law with a mean of 0 and a covariance matrix of... The multivariate normal distribution, which is related to the image A perfect match in dimensions, for example, if For a single-channel image of 32×32 pixels, then Given a 1024-dimensional vector, corresponding to 1024 random variables, these variables all follow a multivariate normal distribution. For a 224×224 pixel three-channel image, then It is a tensor of 3×224×224 dimensions, and the whole follows a multivariate normal distribution.
[0032] Gaussian noise amount to be added The multivariate normal distribution it follows In this context, a mean of 0 can be achieved in The added perturbation is imperceptible to the human eye before and after the image is added. However, the covariance matrix combines noise intensity with the identity matrix to ensure that the noise of each pixel is independent, avoids interference between pixels, and uses noise intensity to control the degree of noise perturbation of each pixel.
[0033] Furthermore, the resulting current noisy image The input is fed into a pre-trained diffusion model to perform stepwise denoising and image reconstruction, resulting in a reconstructed image. The specific process can be represented by Formula 2.
[0034] …………Formula 2 in This represents the diffusion model in the above implementation.
[0035] Finally, the image will be reconstructed. By inputting the data into a target classifier identical to that described above (such as the CIFAR-10 model) and performing the same feature extraction process, the diffusion path feature representation can be obtained. At this point, the dual-path feature extraction is complete.
[0036] In some implementations, please refer to Figure 3 Step 02 specifically includes: Step 021: Based on the feature level, align the direct path feature representation and the diffusion path feature representation so that the direct path feature representation and the diffusion path feature representation with the same feature level correspond to each other. Step 022: Determine the initial feature consistency loss based on the corresponding direct path feature representations and diffusion path feature representations; Step 023: Adjust the initial feature consistency loss function according to the weight coefficients corresponding to different feature levels to determine the feature consistency loss.
[0037] Specifically, after completing dual-path feature extraction and obtaining direct path feature representations and diffusion path feature representations, the next step is to calculate the feature consistency loss based on these two representations to determine the optimization objective of the image purification process. For example, according to the various feature levels of the target classifier, the direct path feature representations and diffusion path feature representations at the same feature level are mapped together. For instance, in the example above, based on the CIFAR-10 model, according to the shallow ( ), middle layer ( ) and deep ( Three levels represent direct path features. Representation of diffusion path characteristics Correspondingly, for example, the direct path features at shallow levels can be represented... Representation of diffusion path characteristics And so on.
[0038] Next, for each feature level, the direct path feature representation is calculated separately. Representation of diffusion path characteristics The similarity between features is calculated using cosine similarity, but Euclidean similarity or MMD similarity can also be used. Then, based on the cosine similarity across all feature levels, an initial feature consistency loss is constructed, as shown in Equation 3.
[0039] …………Formula 3 in For the initial feature consistency loss mentioned above, l For feature level, L A set of feature levels. For feature level l The corresponding direct path feature representation, For feature level l The corresponding diffusion path features are represented as follows: To reconstruct the image, This represents the cosine similarity function. Based on the example above, the feature level in Formula 3... l You can choose block1, block2, or block3.
[0040] Furthermore, considering the varying importance of different feature levels, the initial feature consistency loss needs to be weighted and adjusted based on the weight coefficients of each feature level to differentiate the emphasis of each level. Generally, the weight coefficients increase gradually from shallow to medium to deep layers. For example, in the CIFAR-10 model, the weight coefficients for block1, block2, and block3 are 0.2, 0.3, and 0.5, respectively. The weighted adjustment process for the initial feature consistency loss is detailed in Equation 4, ultimately yielding the feature consistency loss.
[0041] …………Formula 4 in For feature consistency loss, For feature level l The corresponding weighting coefficients. The feature consistency loss described above serves as the optimization objective for the next step of image purification.
[0042] In some implementations, please refer to Figure 4 Step 03 specifically includes: Step 031: Based on the ADAM optimizer, determine the initial image update strategy according to the feature consistency loss, the initial adversarial image or the current adversarial image at the end of the previous iteration, and the preset learning rate, so as to update the original current adversarial image at the end of the current iteration. Step 032: Based on the preset adversarial perturbation amplitude and the initial adversarial image, perform range pruning on the original current adversarial image to determine the current adversarial image at the end of the current iteration.
[0043] Specifically, based on the above implementation method, and given that the feature consistency loss has been determined, the image can be updated based on the current adversarial image or the initial adversarial image at the end of the previous iteration, with the feature consistency loss as the optimization target.
[0044] Regarding the specific update method, in some examples, an initial image update strategy is first determined based on a preset ADAM optimizer, the obtained feature consistency loss, the initial adversarial image or the current adversarial image at the end of the previous iteration, and the learning rate of the already initialized ADAM optimizer. Then, a preliminary cleanup process is performed on the initial adversarial image or the current adversarial image at the end of the previous iteration according to this strategy to obtain the original current adversarial image. Besides using the ADAM optimizer, in memory-constrained scenarios, the SGD optimizer, currently used in related technologies, can also be used instead; the choice can be made based on the specific circumstances.
[0045] The initial image update strategy is as shown in Formula 5.
[0046] ...Formula 5 in Indicates the ADAM optimizer. The current adversarial image at the end of the last iteration. This is the original current adversarial image at the end of the current iteration. For the initial adversarial image, The learning rate for the ADAM optimizer. For gradient operators, This is the feature consistency loss.
[0047] Furthermore, in order to ensure that the size of the optimized image remains within a reasonable range, for example, the image will be further cropped based on the original current adversarial image, as shown in Formula 6.
[0048] …………Formula 6 in, This represents the current adversarial image at the end of the current iteration, where `clip` is the truncation function. The preset adversarial perturbation amplitude is initialized before the image processing method is executed. For example, for the CIFAR-10 model, the adversarial perturbation amplitude is generally initialized to [value missing]. This concludes a single iteration, yielding the current adversarial image at the end of the current iteration. .
[0049] The iteration stopping condition generally includes two parts. At the end of each iteration process, when the current adversarial image at the end of the current iteration is obtained, the following judgment is made: If the number of iterations reaches the pre-initialized maximum number of iterations, then regardless of the processing effect, the iteration is stopped, and the current adversarial image at the end of the last iteration is taken as the final output image.
[0050] If the number of iterations has not reached the pre-initialized maximum number of iterations, but the calculated feature consistency loss is less than the initialized convergence threshold, it indicates that the preset image processing requirements have been met at the current iteration number. In this case, the iteration process can be stopped, and the current adversarial image at the end of the last iteration can be used as the final output image. Thus, the image processing method ends, and the final output image is the cleaned image. Thus, this application establishes a dual-path parallel processing architecture, utilizing the aforementioned dual-path parallel processing in multiple iterations of the image processing process. In the classifier, the consistency difference between natural images and adversarial image samples in the feature space is determined based on the input adversarial image samples. Furthermore, multi-level feature consistency constraints are used to optimize the feature representation, and finally, iterative optimization is performed based on gradient descent to process the input adversarial image samples, thereby achieving purification processing of adversarial image samples. While achieving the processing objective, the stability of the purification process and the full utilization of classifier knowledge are maintained.
[0051] The electronic device in this application includes a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the image processing method described above is implemented.
[0052] The computer-readable storage medium in this application embodiment stores a computer program, which, when executed by a processor, implements the image processing method described above.
[0053] The above description is merely a preferred embodiment of this application and is not intended to limit this application in any way. Although this application has disclosed the preferred embodiment as above, it is not intended to limit this application. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the technical solution of this application. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the technical solution of this application, based on the technical essence of this application and within the spirit and principles of this application, shall still fall within the protection scope of the technical solution of this application.
Claims
1. An image processing method, characterized in that, The method is based on a pre-trained target classifier and a pre-trained diffusion model, and includes: Based on the target classifier and the diffusion model, direct path feature representation and diffusion path feature representation are obtained according to the initial adversarial image or the current adversarial image at the end of the previous iteration. The current adversarial image is determined by updating the initial adversarial image through one or more iterations. The direct path feature representation and the diffusion path feature representation have the same number of feature levels, which are determined by the target classifier. Based on the feature hierarchy, the feature consistency loss is determined according to the direct path feature representation and the diffusion path feature representation; Using the feature consistency loss as an optimization benchmark, an image update strategy is determined to update the current adversarial image at the end of the current iteration.
2. The method according to claim 1, characterized in that, The step of obtaining direct path feature representation and diffusion path feature representation based on the target classifier and the diffusion model, according to the initial adversarial image or the current adversarial image at the end of the previous iteration, includes: Based on the target classifier, the direct path feature representation is determined according to the initial adversarial image or the current adversarial image at the end of the previous iteration; Based on the initial adversarial image or the current adversarial image at the end of the previous iteration, perform Gaussian noise simulation diffusion to determine the current noise image; Based on the diffusion model, denoising and reconstruction are performed according to the current noisy image to determine the reconstructed image; Based on the target classifier, the diffusion path feature representation is determined according to the reconstructed image.
3. The method according to claim 2, characterized in that, The process involves performing Gaussian noise simulation diffusion based on the initial adversarial image or the current adversarial image at the end of the previous iteration to determine the current noise image, using the following formula: in For the current noisy image, The initial adversarial image or the current adversarial image at the end of the previous iteration. The amount of Gaussian noise to be added. The attenuation coefficient is... For noise intensity, It is an identity matrix.
4. The method according to claim 2, characterized in that, The step of determining the feature consistency loss based on the feature hierarchy, according to the direct path feature representation and the diffusion path feature representation, includes: Based on the feature hierarchy, the direct path feature representation and the diffusion path feature representation are aligned so that the direct path feature representation and the diffusion path feature representation with the same feature hierarchy correspond to each other. The initial feature consistency loss is determined based on the corresponding direct path feature representations and diffusion path feature representations. Based on the weight coefficients corresponding to different feature levels, the initial feature consistency loss function is modified to determine the feature consistency loss.
5. The method according to claim 4, characterized in that, The initial feature consistency loss is determined based on the corresponding direct path feature representations and diffusion path feature representations, using the following formula: in The initial feature consistency loss, l For the feature level, L The set of the feature levels, This refers to the direct path feature representation corresponding to the feature level. This represents the diffusion path feature corresponding to the feature level. For the reconstructed image; The initial feature consistency loss function is modified according to the weight coefficients corresponding to different feature levels to determine the feature consistency loss, as shown in the formula: in The feature consistency loss, The weight coefficients are those corresponding to the feature level.
6. The method according to claim 1, characterized in that, The step of determining an image update strategy based on the feature consistency loss to update the current adversarial image at the end of the current iteration includes: Based on the ADAM optimizer, an initial image update strategy is determined according to the feature consistency loss, the initial adversarial image or the current adversarial image at the end of the previous iteration, and a preset learning rate, so as to update the original current adversarial image at the end of the current iteration. Based on the preset adversarial perturbation amplitude and the initial adversarial image, the original current adversarial image is cropped to determine the current adversarial image at the end of the current iteration.
7. The method according to claim 6, characterized in that, The ADAM-based optimizer determines the initial image update strategy based on the feature consistency loss, the initial adversarial image or the current adversarial image at the end of the previous iteration, and a preset learning rate. The formula is as follows: in This refers to the ADAM optimizer. The current adversarial image at the end of the last iteration. This is the original current adversarial image at the end of the current iteration. For the initial adversarial image, The learning rate of the ADAM optimizer. For gradient operators, The feature consistency loss is given.
8. The method according to claim 7, characterized in that, The process involves cropping the original current adversarial image based on a preset adversarial perturbation amplitude and the initial adversarial image to determine the current adversarial image at the end of the current iteration. The formula is as follows: in, This represents the current adversarial image at the end of the current iteration, where `clip` is the truncation function. This is the preset counter-disturbance amplitude.
9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, implements the image processing method as described in any one of claims 1-8.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by the processor, implements the image processing method as described in any one of claims 1-8.