Picture noise reduction method and system based on denoising automatic encoder

Abstract

The invention provides a picture denoising method and system based on a denoising automatic encoder, and the method comprises the steps: obtaining a to-be-processed initial image, carrying out the preprocessing of the initial image, and obtaining a target image; inputting the target image into an image denoising model for image denoising, wherein the image denoising model comprises a feature extraction module, a feature denoising module and a data reconstruction module; wherein the target image is input into the feature extraction module to obtain feature data; the feature data are input into the feature noise reduction module, feature data after noise reduction are obtained, and the feature noise reduction module is composed of a neural memory ordinary differential network; and inputting the feature data after noise reduction into the data reconstruction module to obtain a target image after noise reduction, through the method provided by the invention, noise can be effectively removed, the image can be restored, a novel and low-cost method is provided for image robustness feature extraction and image denoising, and the image processing efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to an image denoising method and system based on a denoising autoencoder. Background Art

[0002] Image denoising is extremely important in the field of image processing. In the process of image acquisition, transmission and storage, it is inevitable that it will be disturbed by various noises, which seriously affects the quality of the image and subsequent processing and analysis. In today's machine learning and computer vision fields, image robust feature extraction is also of great significance. First of all, images in the real world often face many complex changes, such as changes in illumination, different perspectives, and noise interference. Robust feature extraction enables the model to extract key features from the image that are not affected by these factors, ensuring that the image can still be accurately recognized and understood under various harsh conditions. Secondly, for large-scale image data sets, robust features help reduce data redundancy and noise, and improve the training efficiency and accuracy of the model. Furthermore, in practical applications, such as autonomous driving and target recognition, image robust feature extraction is the key to the stable operation of the system. Only by extracting robust features can the system accurately detect targets and recognize scenes in different environments and situations, providing a reliable basis for decision-making.

[0003] In the field of computer vision, methods based on denoising autoencoders (DAE) combined with convolutional layers can reduce feature dimensions, eliminate noise, and extract key vectors. DAE relies on the denoising extension of the autoencoder, which enables the model to try to generate cleaner images when the input data is noisy, thereby enhancing the robustness of the model. In addition, the robustness of the model can also be improved through more complex stacked autoencoders. However, DAE-based methods will fail to a large extent when faced with severe noise, and the complex autoencoder (AE) structure will reduce the efficiency of the model. Therefore, a simple AE structure with strong robustness is needed to meet these challenges. Summary of the invention

[0004] One of the embodiments of this specification provides a method for image denoising based on a denoising autoencoder. The method comprises:

[0005] Acquire an initial image to be processed, and preprocess the initial image to obtain a target image;

[0006] The target image is input into an image denoising model for image denoising, wherein the image denoising model includes a feature extraction module, a feature denoising module and a data reconstruction module; wherein,

[0007] Inputting the target image into the feature extraction module to obtain feature data;

[0008] Inputting the feature data into the feature denoising module to obtain denoised feature data, wherein the feature denoising module is composed of a neural memory ordinary differential network;

[0009] The denoised feature data is input into the data reconstruction module to obtain a denoised target image.

[0010] In some embodiments, the feature extraction module consists of two linear layers;

[0011] The target image is input into a first linear layer to obtain a one-dimensional vector of N elements;

[0012] The one-dimensional vector of N elements is input into the second linear layer and extracted to obtain a one-dimensional vector of M elements, where N≥M.

[0013] In some embodiments, the one-dimensional vector of M elements obtained by the feature extraction layer is input into the ordinary differential equation of the neural memory ordinary differential network:

[0014]

[0015] Among them, x′ represents a one-dimensional vector of M elements, y(t) represents the first activation function, Represents the derivative function of y(t), sin 2 (·) is the internal activation function of the neural memory ordinary differential network;

[0016] The output of the feature denoising module is obtained by solving the neural memory ordinary differential network using the Dormand-Prince method.

[0017] In some embodiments, the data reconstruction module is composed of two linear layers and one activation layer;

[0018] The denoised feature data is passed through two linear layers to obtain reconstructed feature data;

[0019] A denoised target image is obtained through the one activation layer.

[0020] One of the embodiments of this specification provides an image denoising system based on a denoising autoencoder, the system comprising:

[0021] An acquisition module is used to acquire an initial image to be processed, pre-process the initial image, and obtain a target image;

[0022] The image processing module is used to extract the feature data of the target image, obtain the denoised feature data using a neural memory ordinary differential network, and obtain the denoised target image.

[0023] One of the embodiments of the present specification provides an image denoising device based on a denoising auto-encoder, comprising a processor, wherein the processor is used to execute an image denoising method based on a denoising auto-encoder.

[0024] One of the embodiments of the present specification provides a computer-readable storage medium, wherein the storage medium stores computer instructions. When a computer reads the computer instructions in the storage medium, the computer executes an image denoising method based on a denoising autoencoder.

[0025] The beneficial effects of the present invention include:

[0026] The method described in the present invention can extract the robust features of an image without increasing model parameters and complexity, and can also more effectively remove noise and restore the image, thereby providing a novel and low-cost method for image robust feature extraction and image denoising, and improving image processing efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] This specification will be further described in the form of exemplary embodiments, which will be described in detail by the accompanying drawings. These embodiments are not restrictive, and in these embodiments, the same number represents the same structure, wherein:

[0028] Figure 1 is an exemplary flow chart of a method for image denoising based on a denoising autoencoder according to some embodiments of this specification;

[0029] Figure 2 is an exemplary framework diagram of an image denoising model according to some embodiments of this specification;

[0030] Figure 3 is an exemplary structural diagram of a neural constant differential network according to some embodiments of this specification;

[0031] Figure 4 is an exemplary structural diagram of image denoising model training according to some embodiments of this specification;

[0032] Figure 5 is an exemplary structural diagram of an image denoising system based on a denoising autoencoder according to some embodiments of this specification;

[0033] Figure 6 It is a comparison chart of experimental results of the image denoising method based on the denoising autoencoder shown in some embodiments of this specification. DETAILED DESCRIPTION

[0034] In order to more clearly illustrate the technical solutions of the embodiments of this specification, the following is a brief introduction to the drawings required for the description of the embodiments. Obviously, the drawings described below are only some examples or embodiments of this specification. For ordinary technicians in this field, this specification can also be applied to other similar scenarios based on these drawings without creative work. Unless it is obvious from the language environment or otherwise explained, the same reference numerals in the figures represent the same structure or operation.

[0035] It should be understood that the "system", "device", "unit" and / or "module" used herein are a method for distinguishing different components, elements, parts, portions or assemblies at different levels. However, if other words can achieve the same purpose, the words can be replaced by other expressions.

[0036] As shown in this specification and claims, unless the context clearly indicates an exception, the words "a", "an", "an" and / or "the" do not refer to the singular and may also include the plural. Generally speaking, the terms "comprise" and "include" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0037] Flowcharts are used in this specification to illustrate the operations performed by the system according to the embodiments of this specification. It should be understood that the preceding or following operations are not necessarily performed precisely in order. Instead, the steps may be processed in reverse order or simultaneously. At the same time, other operations may also be added to these processes, or one or more operations may be removed from these processes.

[0038] Figure 1 is an exemplary flow chart of a method for image denoising based on a denoising autoencoder according to some embodiments of this specification. The method specifically comprises:

[0039] S1, obtaining an initial image to be processed, preprocessing the initial image to obtain a target image.

[0040] In some embodiments, image data of a public dataset may be used as the initial image to be processed, for example, the handwritten character set MNIST, CIFAR-10 data with 10 target classifications, medical classification images PathMNIST, and satellite image classification images EuroSAT.

[0041] In some embodiments, in order to simulate the interference of different noises on the image, noise addition processing may be performed in the original image to be processed to simulate the interference of noise on the image.

[0042] In some embodiments, the preprocessing may include at least one of an image cropping operation, a contrast adjustment operation, a grayscale conversion operation, and a geometric transformation operation.

[0043] S2, inputting the target image into an image denoising model for image denoising, wherein the image denoising model comprises a feature extraction module, a feature denoising module and a data reconstruction module.

[0044] Step S2 specifically also includes:

[0045] S21, inputting the target image into the feature extraction module to obtain feature data.

[0046] S22 inputs the feature data into the feature denoising module to obtain the denoised feature data, wherein the feature denoising module is composed of a neural memory ordinary differential network.

[0047] S23, inputting the denoised feature data into the data reconstruction module to obtain a denoised target image.

[0048] Figure 2 It is an exemplary framework diagram of the image denoising model shown in some embodiments of this specification.

[0049] In some embodiments, the image denoising model is an autoencoder model constructed based on a neural memory ordinary differential equation network.

[0050] Neural memory ordinary differential equations (nmODE) are a class of neural ordinary differential equations (NODEs) that can realize nonlinear mapping from input to corresponding attractors, and related studies have also proved the robustness brought by this mapping. nmODE has a stable mapping property that can increase the robustness of the model. Secondly, using the nmODE method does not require the application of complex stacking methods during the model optimization process. In addition, the non-homeomorphic nature of nmODE helps to learn useful features more easily.

[0051] In some embodiments, the feature extraction module can be composed of two linear layers, the feature denoising module can be an activation layer composed of a neural memory ordinary differential network, and the feature extraction module and the feature denoising module together constitute an encoder of the image denoising model. The encoder is expressed as shown in formula (1):

[0052] z=f a1 (W1x+b1) (1),

[0053] Where x represents the input target image; W1 represents the weight matrix of the encoder; b1 represents the bias of the encoder; f a1 represents the first activation function, i.e., the ordinary differential equation of the neural memory ordinary differential network, and z represents the denoised feature data, i.e., the output of the feature denoising module.

[0054] Furthermore, the processing of the feature extraction module is as follows:

[0055] The feature extraction module consists of two linear layers;

[0056] The target image is input into a first linear layer to obtain a one-dimensional vector of N elements;

[0057] The one-dimensional vector of N elements is input into the second linear layer and extracted to obtain a one-dimensional vector of M elements, where N ≥ M. As shown in formula (2):

[0058] x′=(W1x+b1) (2),

[0059] Among them, x represents the target image; x′ represents a one-dimensional vector of M elements, that is, the input vector to be activated; W1 represents the weight matrix of the encoder, and b1 represents the bias of the encoder.

[0060] For example, taking MNIST data as an example, the model flattens the image (28x28) into a one-dimensional vector of 784 elements, inputs it to the encoder part of feature extraction, and maps the vector into a one-dimensional vector of 100 elements through the fully connected layer for feature extraction.

[0061] The processing of the feature denoising layer is as follows:

[0062] The one-dimensional vector of M elements obtained by the feature extraction layer is input into the ordinary differential equation of the neural memory ordinary differential network:

[0063]

[0064] Among them, x′ represents a one-dimensional vector of M elements, y(t) represents the solution function, Represents the derivative function of y(t), sin 2 (·) is the internal activation function of the neural memory ordinary differentiable network.

[0065] In some embodiments, according to the method of neural memory ordinary differential network, its ordinary differential equation part is extracted to be used as the activation of the encoder part, and the mapping has an "attractor property". In differential equation theory, attractors are used to describe the behavior of a system converging to a certain state during its evolution. The global attractor can attract all the initial conditions in the system and guide the system to a stable state. Systems with global attractors are not easily disturbed by noise because they tend to remain convergent in the presence of disturbances.

[0066] When the feature denoising module receives an input vector or any input in the neighborhood of the input vector, its stable mapping property can converge them to the same attractor y* (the mapping process is as follows Figure 3 shown), Figure 3 The process of mapping the input x′ to y(T) through the feature denoising module is shown, where y(T) is actually the value of y(t) on T (i.e., y*) in the module (Equation 3), and T is a hyperparameter of the ordinary differential equation (nmFunc) of the neural memory ordinary differential network. The model can be adjusted by T. In this embodiment, T can be set to 1, and y(0) is the initial condition for solving the ODE.

[0067] Specifically, the feature vector obtained by the feature extraction module (i.e., the one-dimensional vector of 100 elements mentioned above) will be input into the neural memory ordinary differential equation network as x′ in formula (3), and an ordinary differential equation to be solved is obtained, and the hyperparameters T and y(0) are set when the model is preset. Then, using y(0) and T, the ordinary differential equation is solved to obtain the vector y(T) as the output of this module. Due to the attractor characteristics of the memorized ordinary differential equation network mentioned above, x′ is received by this module. It should be noted that the dimensions of the vectors y(T) and x′ obtained at this time are the same, as if it is the activation of the input vector x′ of the feature extraction module.

[0068] Furthermore, the output of the feature denoising module is obtained by solving the neural memory ordinary differential network using the Dormand-Prince method.

[0069] In some embodiments, the Dormand-Prince method is a numerical solution method for ordinary differential equations, which can effectively solve the numerical simulation problems of differential equation systems through adaptive step size control and high-order accuracy characteristics.

[0070] The Dormand-Prince method provides accurate and stable results in solving neural memory ordinary differential equations. The advantage of this method is its high-order accuracy and adaptability, which enables more accurate simulation of the dynamic behavior of the neural system for complex image data.

[0071] Through the Dormand-Prince method, we can effectively process neural memory ordinary differential equations and obtain more reliable and accurate solutions, providing a reliable numerical basis for further processing and analysis of the target image.

[0072] The attractor property of the feature denoising module enables the module to obtain a stable output as long as the noise data falls within the neighborhood of the original input, even if the input is subject to some interference. This means that for noisy inputs, the network is able to obtain a robust representation. This illustrates the relationship between the properties of the stable mapping of the feature denoising module and its robustness. This theoretical explanation clarifies the stability principle behind the mapping of the feature denoising module: that is, when the input is locally modified, the stable mapping can ensure that the output remains basically unchanged, thereby achieving noise resistance. The features that have undergone this mapping are called the "robust features" of the image, and an interface is introduced for this robust feature so that it can be received externally for subsequent use.

[0073] In some embodiments, the data reconstruction module is a decoder consisting of two linear layers and one activation layer, and the processing process of the data reconstruction module is as follows:

[0074] The denoised feature data is passed through two linear layers to obtain reconstructed feature data; wherein the first linear layer of the data reconstruction module is used for a one-dimensional vector consisting of M elements, and the second linear layer is used to expand and restore the one-dimensional vector consisting of M elements into a feature vector of N elements.

[0075] A denoised target image is obtained through the one activation layer, wherein the activation layer is composed of a second activation function sigmoid function.

[0076] Specifically, the decoding process of the decoder is expressed as formula (4)

[0077]

[0078] Where W2 represents the weight matrix of the decoder; b2 represents the bias of the encoder; f a2 represents the second activation function, represents the target image after denoising, that is, the output of the data reconstruction module; z represents the feature data after denoising, that is, the output of the feature denoising module.

[0079] In some embodiments, the data reconstruction module is symmetrical with the encoder part. For example, if the encoder has two ordered linear layers of 784 nodes and 300 nodes, then the decoder should symmetrically have two ordered linear layers of 300 nodes and 784 nodes to ensure effective reconstruction of the data. The output of this module also introduces an interface so that it can be received by subsequent tasks.

[0080] Figure 4 is an exemplary structural diagram of image denoising model training according to some embodiments of this specification. Figure 4As shown, the image interfered by noise is input into the model as a training set, and the original image without noise is used as the training target. The above machine learning model is trained to generate an image robust feature extraction and image denoising model.

[0081] The model hyperparameters are adjusted according to different data sets. For example, the input layer and output layer are adjusted accordingly according to the size of the input image, so that the input image can be successfully received by the model. However, the network parameters are the initial settings and untrained initial parameter values. The image denoising model is finally generated by learning and adjusting the parameter values ​​according to the training data.

[0082] The training process includes: first, the noisy image is used as the input image, input into the initial target detection model, and the model operation is performed to obtain the model output result. Then the output result is compared with the original image without noise and the loss is calculated to obtain the model detection error. The network loss function is trained once in each training cycle, and the error is back-propagated using the loss function, and the model parameters are updated iteratively; after a certain number of training cycles, the model training is completed when the model training meets the convergence conditions.

[0083] The specific loss function uses the mean squared error (MSE) to calculate the loss between the model reconstructed image and the original noise-free image. The MSE formula is as follows:

[0084]

[0085] Among them, x i is the training target (i.e., the clean original image), is the model output (i.e. the reconstructed image).

[0086] In particular, in order to enhance the denoising effect of the feature denoising module, a weight threshold limit and an additional loss scheme can also be used, aiming to improve the performance of the feature denoising module by constraining W.

[0087] Among them, the weight threshold limiting scheme will constrain the weight value in W1 during the training process, which will make it easier for the subsequent feature denoising module to apply denoising function to it after obtaining the feature, thereby enhancing the denoising and anti-noise ability of the entire model, as shown in formula (6):

[0088]

[0089] Among them, W1 is the original weight, C is the threshold, and W′1 is the weight after weight constraint.

[0090] In some embodiments, the additional loss scheme will make an adaptive change on the original MSE loss mentioned above. This change is related to the size of W1 and W2 and will change with the change of weights. The loss function will increase with the increase of weights, thereby limiting the size of weights and increasing the denoising performance of the model, as shown in formula (7):

[0091] L=L MSE +αL′ (7),

[0092] in, α is a hyperparameter used to adjust the penalty strength.

[0093] It should be noted that the above description of the above process is only for example and explanation, and does not limit the scope of application of this specification. For those skilled in the art, various modifications and changes can be made to the above process under the guidance of this specification. However, these modifications and changes are still within the scope of this specification.

[0094] Figure 5 is an exemplary structural diagram of an image denoising system based on a denoising autoencoder according to some embodiments of this specification. Figure 5 As shown, the system includes:

[0095] The acquisition module 510 is used to acquire an initial image to be processed, and pre-process the initial image to obtain a target image.

[0096] The image processing module 520 is used to extract the feature data of the target image, obtain the denoised feature data using a neural memory ordinary differential network, and obtain the denoised target image.

[0097] Figure 6 It is a comparison chart of experimental results of the image denoising method based on the denoising autoencoder shown in some embodiments of this specification.

[0098] This embodiment experimentally compares the denoising effect of the method of the present invention with that of the prior art in combination with specific data. The specific experimental process is as follows:

[0099] 1. Give the performance indicators of denoising effect for comparison:

[0100] In order to demonstrate the robustness of the features extracted by the method of the present invention and its impact on the subsequent task results, we used the trained model to generate image features from a test data set with different levels of Gaussian noise (clear, slight noise, medium noise, severe noise). These representations were then classified using the K nearest neighbor classifier algorithm (KNN). KNN is based on a certain distance metric to find the k training samples closest to the test set samples in the training set, and then use the "voting method" to determine the category of the test sample through the found k training samples. Then this distance metric method chooses to use L_1 distance (i.e. Manhattan distance) as its distance metric, and its calculation formula () is as follows:

[0101]

[0102] in, and Indicates the two samples whose distance needs to be calculated. and Represents each feature element in the sample. The classification accuracy of the KNN classifier for the features extracted by the model under different noise conditions is used to evaluate the robustness and effectiveness of the image features extracted by this method.

[0103] Then, the peak signal-to-noise ratio (PSNR) of the reconstructed image and the input image is used to evaluate the model's image reconstruction capability. In addition, the reconstructed images are visually compared to intuitively understand the model's denoising and reconstruction capabilities. The PSNR formula (9) is as follows:

[0104]

[0105] Where MSE stands for Mean Squared Error (MSE), MAX I The maximum value representing the color of an image point.

[0106] Four baseline methods are selected in this paper. In order to cope with the challenge of unstable noise attenuation, they are adjusted to a minimum. In addition, the hyperparameters, training rounds and data of each comparison baseline method are consistent in the experiment. The specific description is as follows:

[0107] 1) DAE: In order to highlight the impact of the improvements of the present invention on noise suppression, the present invention finds it necessary to introduce DAE to compare the algorithm proposed by the present invention. This DAE is an extension of the classic autoencoder, which uses noisy images as input and clean images as training targets while keeping the structure of the autoencoder unchanged.

[0108] 2) DAE-SIREN: DAE-SIREN is a DAE model optimized by a sinusoidal activation function called SIREN, which has been proven to be able to capture image shapes well. Similar to the method of the present invention, it also increases the robustness of the model by enhancing the denoising ability of the activation function.

[0109] 3) VAE: VAE is a generative model that generates and reconstructs data by learning the potential distribution of data. In terms of image denoising, VAE can help identify and remove noise in images by learning the potential feature representation of image data, thereby improving the quality and clarity of the image.

[0110] The parameters are set as follows:

[0111] The method of the present invention sets the following parameters: the learning rate is set to 0.0005, and the number of training rounds is set to 10 for the MNIST dataset because the dataset is relatively simple, and 50 for other datasets because it is found through practice that such parameter settings can make the model reach a fitting state and can effectively avoid overfitting. For the parameters of the model architecture, we ensure that the internal parameters of each model are consistent for the same dataset. For different datasets, the node settings of each layer of the model structure are as follows:

[0112] After training the method of the present invention and other benchmark methods for comparison, we used the above method to evaluate the robustness of the features extracted by them to verify whether each model can still extract robust features under different degrees of noise. The results of the method of the present invention and other benchmark models are compared as follows:

[0113] For the MNIST dataset, the reconstructed images of this method are compared with those of other benchmark methods. Figure 6 , the PSNR indicators of their reconstructed images and the original noise-free images are compared as follows: Noise level DAE DAE-SIREN VAE nBmw Noiseless 53.53 51.67 54.95 52.66 Slight noise 49.88 50.96 49.75 51.16 Moderate noise 46.71 44.70 46.51 49.97 Severe noise 42.76 43.54 44.36 48.30

[0114] To verify that this method is low-cost, we compared the training run time and FLOPs per round with other baseline models to measure the time overhead and model complexity overhead caused by different methods of optimizing the model. The comparison results are as follows:

[0115] The above experimental results can verify that:

[0116] 1) This invention innovatively combines the autoencoder with the neural ordinary differential equation network, proving the feasibility of combining NODE and AE;

[0117] 2) The method proposed in the present invention not only exhibits strong noise immunity, but also has high efficiency;

[0118] 3) In the case of severe noise, the denoised image reconstructed by the method proposed in the present invention is clearer than that by other advanced methods.

[0119] The basic concepts have been described above. Obviously, for those skilled in the art, the above detailed disclosure is only for example and does not constitute a limitation of this specification. Although not explicitly stated here, those skilled in the art may make various modifications, improvements and corrections to this specification. Such modifications, improvements and corrections are suggested in this specification, so such modifications, improvements and corrections still belong to the spirit and scope of the exemplary embodiments of this specification.

[0120] At the same time, this specification uses specific words to describe the embodiments of this specification. For example, "one embodiment", "an embodiment", and / or "some embodiments" refer to a certain feature, structure or characteristic related to at least one embodiment of this specification. Therefore, it should be emphasized and noted that "one embodiment" or "an embodiment" or "an alternative embodiment" mentioned twice or more in different positions in this specification does not necessarily refer to the same embodiment. In addition, certain features, structures or characteristics in one or more embodiments of this specification can be appropriately combined.

[0121] In addition, unless explicitly stated in the claims, the order of the processing elements and sequences described in this specification, the use of alphanumeric characters, or the use of other names are not intended to limit the order of the processes and methods of this specification. Although the above disclosure discusses some invention embodiments that are currently considered useful through various examples, it should be understood that such details are only for illustrative purposes, and the attached claims are not limited to the disclosed embodiments. On the contrary, the claims are intended to cover all modifications and equivalent combinations that are consistent with the essence and scope of the embodiments of this specification. For example, although the system components described above can be implemented by hardware devices, they can also be implemented only by software solutions, such as installing the described system on an existing server or mobile device.

[0122] Similarly, it should be noted that in order to simplify the description disclosed in this specification and thus help understand one or more embodiments of the invention, in the above description of the embodiments of this specification, multiple features are sometimes combined into one embodiment, figure or description thereof. However, this disclosure method does not mean that the features required by the subject matter of this specification are more than the features mentioned in the claims. In fact, the features of the embodiments are less than all the features of the single embodiment disclosed above.

[0123] In some embodiments, numbers describing the number of components and attributes are used. It should be understood that such numbers used in the description of the embodiments are modified by the modifiers "about", "approximately" or "substantially" in some examples. Unless otherwise specified, "about", "approximately" or "substantially" indicate that the numbers are allowed to vary by ±20%. Accordingly, in some embodiments, the numerical parameters used in the specification and claims are approximate values, which may change according to the required features of individual embodiments. In some embodiments, the numerical parameters should take into account the specified significant digits and adopt the general method of retaining digits. Although the numerical domains and parameters used to confirm the breadth of their range in some embodiments of this specification are approximate values, in specific embodiments, the setting of such numerical values ​​is as accurate as possible within the feasible range.

[0124] Each patent, patent application, patent application publication, and other materials, such as articles, books, specifications, publications, documents, etc., cited in this specification are hereby incorporated by reference in their entirety. Except for application history documents that are inconsistent with or conflicting with the contents of this specification, documents that limit the broadest scope of the claims of this specification (currently or later attached to this specification) are also excluded. It should be noted that if the descriptions, definitions, and / or use of terms in the materials attached to this specification are inconsistent or conflicting with the contents described in this specification, the descriptions, definitions, and / or use of terms in this specification shall prevail.

[0125] Finally, it should be understood that the embodiments described in this specification are only used to illustrate the principles of the embodiments of this specification. Other variations may also fall within the scope of this specification. Therefore, as an example and not a limitation, alternative configurations of the embodiments of this specification may be considered consistent with the teachings of this specification. Accordingly, the embodiments of this specification are not limited to the embodiments explicitly introduced and described in this specification.

Claims

1. A method for image denoising based on a denoising autoencoder, characterized in that: include: Acquire an initial image to be processed, and preprocess the initial image to obtain a target image; The target image is input into an image denoising model for image denoising, wherein the image denoising model includes a feature extraction module, a feature denoising module and a data reconstruction module; wherein, Inputting the target image into the feature extraction module to obtain feature data; Inputting the feature data into the feature denoising module to obtain denoised feature data, wherein the feature denoising module is composed of a neural memory ordinary differential network; The denoised feature data is input into the data reconstruction module to obtain a denoised target image.

2. The image denoising method based on denoising autoencoder according to claim 1, characterized in that: include: The feature extraction module consists of two linear layers; The target image is input into a first linear layer to obtain a one-dimensional vector of N elements; The one-dimensional vector of N elements is input into the second linear layer and extracted to obtain a one-dimensional vector of M elements, where N≥M.

3. The image denoising method based on denoising autoencoder according to claim 2, characterized in that: include: The one-dimensional vector of M elements obtained by the feature extraction layer is input into the ordinary differential equation of the neural memory ordinary differential network: Among them, x′ represents a one-dimensional vector of M elements, y(t) represents the solution function, Represents the derivative function of y(t), sin 2 (·) is the internal activation function of the neural memory ordinary differential network; The output of the feature denoising module is obtained by solving the neural memory ordinary differential network using the Dormand-Prince method.

4. The image denoising method based on denoising autoencoder according to claim 3, characterized in that: include: The data reconstruction module is composed of two linear layers and one activation layer; The denoised feature data is passed through two linear layers to obtain reconstructed feature data; A denoised target image is obtained through the one activation layer.

5. A picture denoising system based on a denoising autoencoder, characterized in that: include: An acquisition module is used to acquire an initial image to be processed, pre-process the initial image, and obtain a target image; The image processing module is used to extract the feature data of the target image, obtain the denoised feature data using a neural memory ordinary differential network, and obtain the denoised target image.

6. A picture denoising device based on a denoising auto-encoder, comprising a processor, wherein the processor is used to execute the picture denoising method based on a denoising auto-encoder according to any one of claims 1 to 4.

7. A computer-readable storage medium, wherein the storage medium stores computer instructions. When a computer reads the computer instructions in the storage medium, the computer executes the image denoising method based on a denoising autoencoder as claimed in any one of claims 1 to 4.

Images