Image noise reduction method and device, electronic equipment and storage medium

Abstract

The embodiment of the invention discloses an image noise reduction method and device, electronic equipment and a storage medium, and the method comprises the steps: obtaining an image to be subjected to noise reduction, carrying out the frequency domain mapping and noise reduction of the image to be subjected to noise reduction, and obtaining initial noise reduction sub-band images of at least two first frequency domain segments; target noise reduction sub-band images corresponding to the initial noise reduction sub-band images are determined, target noise reduction sub-band images of at least two first frequency domain segments are obtained, the target noise reduction sub-band images are obtained through N times of frequency domain mapping and noise reduction processing on the basis of the initial noise reduction sub-band images, N is larger than or equal to 0, and N is an integer; based on the target noise reduction sub-band images of the at least two first frequency domain segments, performing fusion processing and spatial domain mapping to obtain a target noise reduction image; thus, the frequency domain features of the image to be denoised are fully utilized, deep exploration of sub-bands of different frequency bands and information aggregation across the frequency bands are achieved, frequency domain and space domain information is fully utilized, high-frequency details of the image are reserved as much as possible while noise is reduced, and a good noise reduction effect is achieved.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and in particular to an image noise reduction method, device, electronic device and storage medium. Background Art

[0002] Due to the limitations of shooting environment, hardware and software conditions of image acquisition equipment, etc., images captured by image acquisition equipment often contain certain noise, resulting in reduced image quality. Therefore, how to reduce or even remove noise in images is crucial. Summary of the invention

[0003] The embodiments of the present application are intended to provide an image noise reduction method, device, electronic device and storage medium to solve the problem of poor noise reduction effect that may exist in the related art.

[0004] The technical solution of this application is implemented as follows:

[0005] In a first aspect, an embodiment of the present application provides an image noise reduction method, the method comprising:

[0006] Obtaining an image to be denoised, performing frequency domain mapping and denoising processing on the image to be denoised, and obtaining at least two initial denoised sub-band images of a first frequency domain segment;

[0007] Determine a target denoised subband image corresponding to each of the initial denoised subband images, and obtain at least two target denoised subband images of the first frequency domain segment, wherein the target denoised subband image is obtained by performing frequency domain mapping and denoising processing N times based on the initial denoised subband image, where N≥0, and N is an integer;

[0008] Based on the target denoised sub-band images of the at least two first frequency domain segments, fusion processing and spatial domain mapping are performed to obtain a target denoised image.

[0009] In a second aspect, an embodiment of the present application provides an image noise reduction device, the device comprising:

[0010] An acquisition module, used for acquiring an image to be denoised;

[0011] A processing module, configured to perform frequency domain mapping and noise reduction processing on the image to be denoised, to obtain at least two initial denoised sub-band images of the first frequency domain segment;

[0012] A determination module, configured to determine a target denoised subband image corresponding to each of the initial denoised subband images, and obtain at least two target denoised subband images of the first frequency domain segment, wherein the target denoised subband image is obtained by performing N frequency domain mapping and denoising processing on the initial denoised subband image, where N≥0, and N is an integer;

[0013] The processing module is further used to perform fusion processing and spatial domain mapping based on the target denoised sub-band images of the at least two first frequency domain segments to obtain a target denoised image.

[0014] In a third aspect, an embodiment of the present application provides an electronic device, the electronic device comprising: a memory and a processor;

[0015] The memory stores a computer program executable on the processor.

[0016] When the processor executes the computer program, part or all of the steps in the method described in the first aspect are implemented.

[0017] In a fourth aspect, an embodiment of the present application provides a storage medium, characterized in that the storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement part or all of the steps in the method described in the first aspect.

[0018] In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program or instructions, which, when executed by a processor, implements some or all of the steps in the method described in the first aspect.

[0019] The embodiments of the present application provide an image denoising method, device, electronic device and storage medium, which obtain an image to be denoised, perform frequency domain mapping and denoising processing on the image to be denoised, and obtain at least two initial denoised subband images of the first frequency domain segments; determine the target denoised subband images corresponding to each initial denoised subband image, and obtain at least two target denoised subband images of the first frequency domain segments, wherein the target denoised subband image is obtained by performing N frequency domain mapping and denoising processing on the initial denoised subband image, N≥0, and N is an integer; based on the target denoised subband images of the at least two first frequency domain segments, fusion processing and spatial domain mapping are performed to obtain the target denoised image; in this way, the embodiments of the present application make full use of the frequency domain characteristics of the input image to be denoised, have both in-depth exploration of subbands of different frequency bands and information aggregation across frequency bands, make full use of frequency domain and spatial domain information, and can retain the high-frequency details of the image as much as possible while denoising, thereby achieving a good denoising effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 An optional flowchart of the image noise reduction method provided in the embodiment of the present application;

[0021] Figure 2 A schematic diagram of a network structure of an image denoising method provided in an embodiment of the present application;

[0022] Figure 3 An optional structural diagram of a cross-band information aggregation module provided in an embodiment of the present application;

[0023] Figure 4A An optional structural diagram of a noise reduction network corresponding to a high frequency domain provided in an embodiment of the present application;

[0024] Figure 4B An optional structural diagram of a noise reduction network corresponding to a low-frequency domain segment provided in an embodiment of the present application;

[0025] Figure 5 An optional flowchart of the image noise reduction method provided in the embodiment of the present application;

[0026] Figure 6 An optional structural diagram of a noise reduction network provided in an embodiment of the present application;

[0027] Figure 7 An optional flowchart of the image noise reduction method provided in the embodiment of the present application;

[0028] Figure 8 A schematic diagram of a network structure of an image denoising method provided in an embodiment of the present application;

[0029] Fig. 9 An optional flowchart of the image noise reduction method provided in the embodiment of the present application;

[0030] Fig.10 An optional flowchart of the image noise reduction method provided in the embodiment of the present application;

[0031] Fig.11 An optional structural diagram of an image noise reduction device provided in an embodiment of the present application;

[0032] Fig.12 An optional structural diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0033] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0034] It should be understood that the "embodiments of the present application" or "the aforementioned embodiments" mentioned throughout the specification mean that specific features, structures or characteristics related to the embodiments are included in at least one embodiment of the present application. Therefore, "in the embodiments of the present application" or "in the aforementioned embodiments" appearing throughout the specification may not necessarily refer to the same embodiment. In addition, these specific features, structures or characteristics can be combined in one or more embodiments in any suitable manner. In the various embodiments of the present application, the size of the sequence number of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application. The above-mentioned sequence numbers of the embodiments of the present application are for description only and do not represent the advantages and disadvantages of the embodiments.

[0035] For practical needs, users need to use image acquisition devices, such as mobile phones and cameras, to capture images of the current scene. However, due to the noise signals in the environment and / or the limitations of the hardware and software of the image acquisition devices, there is a certain amount of noise in the captured images, resulting in the captured images not meeting the actual needs of users. Therefore, how to reduce or even remove the noise in the image is crucial.

[0036] In the related technologies, image denoising schemes that combine frequency domain information are usually divided into schemes based on traditional algorithms and schemes based on deep learning. The schemes based on traditional algorithms essentially map the image to the frequency domain, filter and denoise the components of the image in each frequency band using traditional schemes, and calculate the fusion coefficients. When performing frequency domain inverse mapping, these coefficients are applied to combine the filtered components of different frequency bands together to achieve image denoising; while most schemes based on deep learning also use frequency domain transformation and inverse transformation as means of image downsampling and upsampling, respectively, to replace the original down / upsampling operators, and the rest of the calculations are still the commonly used convolution, attention mechanism and other calculation modules in the deep learning model.

[0037] Specifically, in one solution of the related technology, the image to be denoised is first mapped to the frequency domain, and then the obtained sub-band images of multiple different frequency bands are merged in the channel dimension; then, the merged multi-channel image is input into the neural network for denoising, and finally the denoised multi-channel image is mapped back to the spatial domain to obtain the denoised image. However, there are differences in the details of the sub-band images of different frequency bands in the image. This method merges the channels of the frequency band components mapped to the frequency domain and sends them to the neural network for processing, which may result in poor denoising effect.

[0038] In another scheme of the related technology, the target original RAW image to be denoised is preprocessed with a pre-acquired noise model to obtain a noise component image; the target RAW image and the noise component image are input into a pre-trained AI denoising model to obtain a denoised RAW image; wherein the AI ​​denoising network model includes: a deeply separable channel attention module, an inverted residual channel attention module, a short-circuited inverted residual channel attention module, a discrete wavelet transform module, and an inverse discrete wavelet transform module. It can be seen that the AI ​​denoising model uses discrete wavelet transform and inverse discrete wavelet transform as means of downsampling and upsampling, respectively, and the rest of the network structure is designed as a computing module such as the commonly used convolution and attention mechanism in the network structure based on deep learning; at the same time, this method is limited by the image properties of the image, that is, the image is a RAW domain image, and the noise component image needs to be obtained.

[0039] In view of one or more of the above problems, an exemplary embodiment of the present application first provides an image denoising method. When performing image processing, the image to be denoised can be divided into frequency domain segments and denoised. Based on the initial denoised subband images of different first frequency domain segments, zero, one or more frequency domain mappings and denoising processes are performed to obtain corresponding target denoised subband images. After that, the target denoised subband images of different first frequency domain segments are fused and spatially mapped. In this way, the frequency domain features of the input image to be denoised can be fully utilized, the initial denoised subband images of different frequency domain segments are deeply explored, and the cross-band information between the target denoised subband images of different frequency domain segments is aggregated, thereby achieving efficient and high-quality image denoising processing, and thus achieving better denoising effect.

[0040] Reference Figure 1 , Figure 1 An image noise reduction method provided in an embodiment of the present application can be applied to any electronic device with image noise reduction capability. The method comprises the following steps:

[0041] Step 101: Obtain an image to be denoised, perform frequency domain mapping and denoising on the image to be denoised, and obtain at least two initial denoised sub-band images of a first frequency domain segment.

[0042] In the embodiment of the present application, the image to be denoised may be an image captured by the execution subject of the image denoising method provided by the present application, or an image sent by other electronic devices received by the execution subject, or an image read by the execution subject from a preset storage medium. The image to be denoised may be a complete image or a sub-image in a certain area of ​​the complete image.

[0043] For example, taking a mobile phone as the execution subject, assuming that the mobile phone captures an image with noise, the image can be used as the image to be denoised, or a sub-image in a specific area of ​​the image can be used as the image to be denoised. The specific area can be a preset area or an area determined according to the operation instructions input by the user. This application does not impose any restrictions on this.

[0044] In the embodiment of the present application, the image to be denoised may be a single-frame image, or multiple frames of images that are not related to each other, or multiple frames of images that are related to each other, such as multiple consecutive frames of images of a video, which is not limited in the present application. In some embodiments, the image to be denoised may be a spatial domain image, or may be an image that has been processed by a spatial domain processing method, such as an image that has been linearly stretched or an image that has been enlarged by an integer multiple of the image.

[0045] In an embodiment of the present application, the first frequency domain segment is a frequency domain segment obtained by dividing the frequency domain segment in which the image to be denoised is located, and at least two first frequency domain segments may be a predetermined first number of preset frequency domain segments; at least two first frequency domain segments may also be a target number of frequency domain segments determined based on the pixels of the image to be denoised.

[0046] In an achievable scenario, at least two first frequency domain segments are a predetermined first number of preset frequency domain segments. As an example, assuming that the first number can be 2, the first number of preset frequency domain segments can include a high frequency segment and a low frequency segment; after obtaining the image to be denoised, each image to be denoised is divided into sub-band images corresponding to the two frequency domain segments, namely, the high frequency segment and the low frequency segment. As an example, assuming that the first number can be 4, the first number of preset frequency domain segments can include a low frequency segment, a horizontal high frequency segment, a vertical high frequency segment, and a diagonal high frequency segment; after obtaining the image to be denoised, each image to be denoised is divided into sub-band images corresponding to the four frequency domain segments, namely, the low frequency segment, the horizontal high frequency segment, the vertical high frequency segment, and the diagonal high frequency segment.

[0047] In another possible scenario, at least two first frequency domain segments are frequency domain segments of a target number determined based on pixels of the image to be denoised. As an example, the corresponding target number is determined according to the specific situation of the image to be denoised by a neural network for frequency domain segmentation; it should be noted that the neural network for frequency domain segmentation has the function of converting a spatial domain image into a frequency domain image.

[0048] For example, after the image 1 to be denoised is input into the neural network for frequency domain segmentation, the image 1 to be denoised is divided into three frequency domain segments, namely, a high frequency segment, a medium frequency segment, and a low frequency segment; after the image 2 to be denoised is input into the neural network for frequency domain segmentation, the image 2 to be denoised is divided into five frequency domain segments, namely, a high frequency segment, a sub-high frequency segment, a medium frequency segment, a sub-low frequency segment, and a low frequency segment; after the image 3 to be denoised is input into the neural network for frequency domain segmentation, the image 3 to be denoised is divided into four frequency domain segments, namely, a low frequency segment, a horizontal high frequency segment, a vertical high frequency segment, and a diagonal high frequency segment.

[0049] In the embodiment of the present application, the initial denoised sub-band image of the first frequency domain segment can be understood as a denoised sub-band image obtained after the image to be denoised is subjected to one frequency domain mapping and one denoising process.

[0050] After obtaining the image to be denoised, the electronic device can convert the image to be denoised into a frequency domain image, divide the frequency domain segment based on the obtained frequency domain image, obtain at least two sub-band images to be denoised in the first frequency domain segment, and perform noise reduction processing on the sub-band images to be denoised in the at least two first frequency domain segments respectively to obtain at least two initial denoised sub-band images in the first frequency domain segment. It should be noted that the frequency domain mapping and the noise reduction processing correspond one to one, that is, after the image to be denoised is subjected to one or more frequency domain mappings, one or more of the sub-band images of the at least two frequency domain segments obtained after each frequency domain mapping will be subjected to a noise reduction processing.

[0051] Step 102: determine the target denoised sub-band images corresponding to the initial denoised sub-band images, and obtain at least two target denoised sub-band images in the first frequency domain segment.

[0052] The target denoised sub-band image is obtained by performing N frequency domain mapping and denoising processes on the initial denoised sub-band image, where N≥0 and N is an integer.

[0053] In the embodiment of the present application, the target denoised sub-band image can be an image obtained by performing zero, one or multiple frequency domain mapping and denoising processes based on the initial denoised sub-band image. For example, for a sub-band image rich in edge and texture information in a high frequency band, multiple frequency domain mapping and denoising processes can more accurately retain these key details while removing noise; for a sub-band image representing a smooth area of ​​the image in a low frequency band, noise can also be suppressed in a targeted manner to maintain the smoothness of the area.

[0054] In some embodiments, N corresponding to different first frequency domain segments are the same or different, and N corresponding to at least one first frequency domain segment is greater than a times threshold.

[0055] In the embodiment of the present application, the number threshold can be 0.

[0056] In the embodiment of the present application, the value of N is related to the performance of the device running the network, including but not limited to power consumption limit and denoising effect. For example, if the computing power of the device running the network is relatively small, and there are power consumption and operation delay limitations, then N can be set smaller, such as N=1 for all frequency domain segments, or N=1 for the low-frequency domain segment, and N=0 for the remaining frequency bands; if after testing, the benefits of increasing N are not obvious, then N can be maintained at a relatively small value. It should be noted that the embodiments of the present application are only examples and are not specifically limited.

[0057] It can be understood that the times of re-frequency domain mapping and denoising processing for the initial denoised subband images of different first frequency domain segments are the same or different, and the frequency domain mapping times of at least one image of the first frequency domain segment is greater than zero.

[0058] Here, an example is taken for explanation that at least two first frequency domain segments are four frequency domain segments, namely, a low frequency domain segment, a horizontal high frequency domain segment, a vertical high frequency domain segment, and a diagonal high frequency domain segment, which may include various situations.

[0059] Case 1: The N corresponding to the horizontal high-frequency domain segment, the vertical high-frequency domain segment, and the diagonal high-frequency domain segment is 0, and the N corresponding to the low-frequency domain segment is 1;

[0060] Case 2: The corresponding N of the low-frequency domain segment, the horizontal high-frequency domain segment, the vertical high-frequency domain segment and the diagonal high-frequency domain segment can be any one of 1, 2, 3, etc. The value of N can be set in advance. Of course, N can also be set to other numbers besides the above examples.

[0061] Case three: Based on the image information (such as image details, image content or image texture, etc.) of the sub-band image to be denoised corresponding to different first frequency domain segments, the N value corresponding to the image information is set. Here, the image information can be obtained based on the pixels in the sub-band image to be denoised.

[0062] Case 4: one or part of the low-frequency domain segment, horizontal high-frequency domain segment, vertical high-frequency domain segment and diagonal high-frequency domain segment, such as the low-frequency domain segment, horizontal high-frequency domain segment, vertical high-frequency domain segment, the corresponding N can be any one of 1, 2, 3, etc., and the N in the remaining frequency domain segments such as the diagonal high-frequency domain segment can be set to 0. The value of N can be set in advance. Of course, N can also be set to other numbers other than the above examples.

[0063] It should be noted that the above examples are only illustrative. The embodiments of the present application can also be set according to other situations. It only needs to meet the conditions that the N corresponding to different first frequency domain segments are the same or different, and the N corresponding to at least one first frequency domain segment is greater than the number threshold.

[0064] In an embodiment of the present application, after the electronic device performs frequency domain mapping and noise reduction processing on the obtained image to be denoised to obtain at least two initial noise reduction sub-band images of the first frequency domain segments, the target noise reduction sub-band image corresponding to each initial noise reduction sub-band image of the first frequency domain segment is determined in the initial noise reduction sub-band images of the at least two first frequency domain segments, thereby obtaining the target noise reduction sub-band images of the at least two first frequency domain segments. In this way, the embodiment of the present application performs zero, one or multiple frequency domain mapping and noise reduction processing on each initial noise reduction sub-band image, and can gradually optimize the noise reduction effect in the multiple processing processes according to the characteristics of the sub-band images of different frequency bands, better adapt to the details of different frequency bands, thereby effectively improving the overall noise reduction quality; at the same time, after N times of frequency domain mapping and noise reduction processing, each processing can optimize the result of the previous time and gradually reduce the noise residue. As the number of processing times increases, the noise reduction sub-band image is closer to the ideal noise-free state, and compared with only one processing, it can remove noise more deeply and significantly improve the noise reduction effect. Finally, since each initial denoised sub-band image is processed independently and multiple times, the embodiment of the present application can retain the detail information of each frequency band of the image more finely during the denoising process. In multiple frequency domain mapping and denoising processes, appropriate processing methods can be used according to the characteristics of different frequency bands to avoid the loss of details caused by unified processing. Fusion processing of target denoised sub-band images of different frequency bands can integrate the effective detail information retained after fine processing of each frequency band, give full play to the advantages of each frequency band, and further enhance the overall detail expression of the image. This fusion method ensures that while reducing noise, various detail information of the image can be retained and optimized, thereby improving the visual quality of the image.

[0065] Step 103: Based on the target denoised subband images of at least two first frequency domain segments, fusion processing and spatial domain mapping are performed to obtain a target denoised image.

[0066] In the embodiment of the present application, the target denoised image is an image obtained after denoising the image to be denoised. The target denoised image can be understood as an image obtained after fusion processing and spatial domain mapping based on at least two target denoised subband images of the first frequency domain segment in different fusion modes.

[0067] In an embodiment of the present application, after obtaining the target denoised subband images of at least two first frequency domain segments, the electronic device performs multiple different fusion processing and one spatial domain mapping based on the target denoised subband images of at least two first frequency domain segments to obtain the target denoised image.

[0068] An image denoising method provided in an embodiment of the present application obtains an image to be denoised, performs frequency domain mapping and denoising processing on the image to be denoised, and obtains at least two initial denoised subband images of the first frequency domain segments; determines a target denoised subband image corresponding to each initial denoised subband image, and obtains at least two target denoised subband images of the first frequency domain segments, wherein the target denoised subband image is obtained by performing N frequency domain mapping and denoising processing on the initial denoised subband image, N≥0, and N is an integer; based on the target denoised subband images of the at least two first frequency domain segments, fusion processing and spatial domain mapping are performed to obtain the target denoised image; in this way, the embodiment of the present application fully utilizes the frequency domain characteristics of the input image to be denoised, has both in-depth exploration of subbands of different frequency bands and information aggregation across frequency bands, and fully utilizes the frequency domain and spatial domain information, and can retain the high-frequency details of the image as much as possible while denoising, thereby achieving a good denoising effect.

[0069] In some embodiments, step 101 obtains the image to be denoised, performs frequency domain mapping and denoising on the image to be denoised, and obtains initial denoised sub-band images of at least two first frequency domain segments, which can be achieved through the following steps.

[0070] Step A1: Perform frequency domain mapping on the image to be denoised to obtain at least two sub-band images to be denoised in the first frequency domain segment.

[0071] In the embodiment of the present application, the first frequency domain segment is a frequency domain segment obtained by dividing the frequency domain segment where the image to be denoised is located. The determination of the number of the first frequency domain segment has been described in the above embodiment and will not be repeated here. The sub-band image to be denoised is used to represent image information of different frequency domain segments in the image to be denoised.

[0072] In the embodiment of the present application, the frequency domain mapping of the image to be denoised can be performed by wavelet transform, such as discrete wavelet transform (DWT) or Haar wavelet transform (HWT), to map the image to be denoised to the frequency domain. Of course, the image to be denoised can also be mapped to the frequency domain by Fourier transform. This application does not make any specific restrictions on this.

[0073] In one possible scenario, refer to Figure 2 As shown, the input image to be denoised is mapped to the frequency domain through the DWT module 201, and the sub-band image to be denoised mapped in the frequency domain is split through the channel splitting (SPLIT) module 202 to obtain four sub-band images to be denoised in the first frequency domain segments, namely, the sub-band image to be denoised Z1 in the low frequency domain segment, the sub-band image to be denoised Z2 in the horizontal high frequency domain segment, the sub-band image to be denoised Z3 in the vertical high frequency domain segment, and the sub-band image to be denoised Z4 in the diagonal high frequency domain segment.

[0074] Step A2: performing denoising processing on at least two sub-band images to be denoised in the first frequency domain segments respectively to obtain at least two initial denoised sub-band images in the first frequency domain segments.

[0075] In an embodiment of the present application, denoising processing of sub-band images of each frequency domain segment can be achieved through a first denoising network, different frequency domain segments can correspond to different first denoising networks, and each frequency domain segment can correspond to a first denoising network, wherein the denoising network is used to perform denoising processing on sub-band images of each frequency domain segment.

[0076] It can be understood that each frequency domain segment can be a first frequency domain segment. Of course, each frequency domain segment can also be a divided frequency domain segment obtained after the image to be denoised undergoes one, two or more frequency domain mappings.

[0077] In some embodiments, the denoising network includes a convolutional neural network, or a neural network of different depths including a cross-band information aggregation module, and the denoising network includes a first denoising network.

[0078] In the embodiment of the present application, the noise reduction network can be a traditional convolutional neural network (CNN). Of course, the noise reduction network can also be a neural network of different depths including a cross-band information aggregation module (CBIA). Among them, the cross-band information aggregation module is a neural network used to extract frequency domain features of each frequency band sub-band image. At the same time, the cross-band information aggregation module is a neural network that combines spatial domain convolution and frequency domain convolution. Exemplarily, the network structure of the cross-band information aggregation module is used to further extract and aggregate frequency domain information, such as Figure 3 As shown, the input image input to the cross-band information aggregation module is subjected to information extraction and interaction through the first convolution module Conv2D, and then frequency domain and spatial domain splitting is performed through the channel splitting (SPLIT) module. The frequency domain features in the input image are extracted through the Fast Fourier Transform (FFT) module. After the second convolution module Conv2D aggregates information in the frequency domain, the Inverse Fast Fourier Transform (IFFT) module restores the aggregated input feature image obtained after frequency domain aggregation to the original data domain, and concatenates and fuses the feature image obtained by the third convolution module Conv2D, i.e., pure spatial domain convolution, through the fourth convolution module Conv2D to exchange information and obtain the output image. In this way, after the module containing FFT, the denoising and detail restoration effects of the network on repeated dense textures are improved.

[0079] It should be noted that the network structure of the CBIA module includes but is not limited to Figure 3 The network structure shown in the figure, the network structure of the CBIA module can be any structure that can aggregate the spatial domain and the frequency domain, or it can include Figure 3 Various permutations and combinations of up-convolution module, FFT module, and IFFT module.

[0080] Different frequency domain segments may correspond to different denoising networks, and each frequency domain segment may correspond to a denoising network. In a feasible scenario, it is assumed that the image to be processed is divided into two sub-band images to be denoised corresponding to the first frequency domain segment, namely, the high-frequency domain segment and the low-frequency domain segment. For the sub-band images to be denoised in the high-frequency domain segment, the information content is relatively small and may not require much processing. Therefore, the denoising network corresponding to the high-frequency domain segment may be a network structure that does not perform downsampling, such as Figure 4A As shown in , the denoising network corresponding to the high-frequency domain segment can include two convolution modules and one CBIA module; for the sub-band image to be denoised in the low-frequency domain segment, the information content is relatively rich, and the denoising network corresponding to the low-frequency domain segment can adopt a more complex network structure, such as Figure 4B As shown, the corresponding noise reduction network of the low frequency domain segment may include at least two convolution modules, one or more CBIA modules, a down-sampling (Down Scale) module and an up-sampling (Up Scale) module, etc.

[0081] In a feasible scenario, continue to refer to Figure 2 The four sub-band images to be denoised in the first frequency domain segments split by the SPLIT module 202 are respectively subjected to denoising processing by corresponding denoising networks to obtain initial denoised sub-band images corresponding to the first frequency domain segments. Optionally, the subband image Z1 to be denoised in the low-frequency domain segment is subjected to denoising processing by a first denoising network such as FUNET network 2031 to obtain an initial denoised subband image Z'1 of the low-frequency domain segment; the subband image Z2 to be denoised in the horizontal high-frequency domain segment is subjected to denoising processing by a first denoising network such as FUNET network 2032 to obtain an initial denoised subband image Z'2 of the horizontal high-frequency domain segment; the subband image Z3 to be denoised in the vertical high-frequency domain segment is subjected to denoising processing by a first denoising network such as FUNET network 2033 to obtain an initial denoised subband image Z'3 of the vertical high-frequency domain segment; the subband image Z4 to be denoised in the diagonal high-frequency domain segment is subjected to denoising processing by a first denoising network such as FUNET network 2034 to obtain an initial denoised subband image Z'4 of the diagonal high-frequency domain segment. In this way, a denoised image is obtained after denoising each subband image once.

[0082] In some embodiments, determining the target denoised sub-band images corresponding to the initial denoised sub-band images in step 102 may be achieved through one or more of the following steps B1, B2 and B3.

[0083] Step B1: If N is equal to 0, the initial denoised sub-band image corresponding to the first frequency domain segment where N is equal to 0 is determined as the target denoised sub-band image corresponding to the first frequency domain segment where N is equal to 0.

[0084] In the embodiment of the present application, N equals 0, indicating that the initial denoised subband image corresponding to the first frequency domain segment will not be subjected to frequency domain mapping and denoising processing after the aforementioned one-time frequency domain mapping and denoising processing. At this time, the initial denoised subband image corresponding to the first frequency domain segment can be used as the target denoised subband image corresponding to the first frequency domain segment.

[0085] In a feasible scenario, continue to refer to Figure 2 ,Depend on Figure 2 It can be seen that the initial denoised subband image Z'2 of the horizontal high-frequency domain segment, the initial denoised subband image Z'3 of the vertical high-frequency domain segment, and the initial denoised subband image Z'4 of the diagonal high-frequency domain segment will no longer be subjected to frequency domain mapping and denoising processing. At this time, the initial denoised subband image Z'2 of the horizontal high-frequency domain segment can be used as the target denoised subband image Z"2 of the horizontal high-frequency domain segment, the initial denoised subband image Z'3 of the vertical high-frequency domain segment can be used as the target denoised subband image Z"3 of the vertical high-frequency domain segment, and the initial denoised subband image Z'4 of the diagonal high-frequency domain segment can be used as the target denoised subband image Z"4 of the diagonal high-frequency domain segment. In this way, the target denoised subband images corresponding to some of the first frequency domain segments in at least two first frequency domain segments are obtained.

[0086] Step B2: If N is equal to 1, based on the initial denoised subband image corresponding to the first frequency domain segment where N is equal to 1, perform frequency domain mapping and denoising processing to obtain at least two intermediate denoised subband images of the second frequency domain segments; based on the intermediate denoised subband images of the at least two second frequency domain segments and the initial denoised subband image corresponding to the first frequency domain segment where N is equal to 1, obtain the target denoised subband image corresponding to the first frequency domain segment where N is equal to 1, wherein the first frequency domain segment where N is equal to 1 includes at least two second frequency domain segments.

[0087] In the embodiment of the present application, N is equal to 1, which indicates that after the initial denoised sub-band image corresponding to the first frequency domain segment has undergone the aforementioned frequency domain mapping and denoising processing, it is necessary to perform another frequency domain mapping and denoising processing subsequently, that is, to perform an in-depth exploration of the sub-band image corresponding to the first frequency domain segment.

[0088] In the embodiment of the present application, the intermediate denoised sub-band image is an image obtained after performing frequency domain mapping and denoising processing on the initial denoised sub-band image.

[0089] In the embodiment of the present application, the second frequency domain segment is a frequency domain segment obtained by dividing the first frequency domain segment with N equal to 1, and at least two second frequency domain segments may be a predetermined first number of preset frequency domain segments; at least two second frequency domain segments may also be a target number of frequency domain segments determined based on pixels of the initial denoised subband image of the first frequency domain segment with N equal to 1. The method for determining the number of the second frequency domain segments is similar to the method for determining the number of the first frequency domain segments, and will not be repeated here.

[0090] In an embodiment of the present application, after the electronic device determines the initial denoised subband image corresponding to the first frequency domain segment where N is equal to 1, the initial denoised subband image corresponding to the first frequency domain segment where N is equal to 1 is subjected to frequency domain mapping and denoising processing again to obtain at least two intermediate denoised subband images of the second frequency domain segments; further, based on the intermediate denoised subband images of the at least two second frequency domain segments and the initial denoised subband image corresponding to the first frequency domain segment where N is equal to 1, a target denoised subband image corresponding to the first frequency domain segment where N is equal to 1 is obtained, wherein the first frequency domain segment where N is equal to 1 includes at least two second frequency domain segments.

[0091] In a feasible scenario, continue to refer to Figure 2 ,Depend on Figure 2 It can be seen that the initial denoised subband image Z'1 of the low frequency domain segment needs to be mapped in the frequency domain and denoised again later, that is, the subband image corresponding to the first frequency domain segment is deeply explored. At this time, the initial denoised subband image Z'1 of the low frequency domain segment is mapped to the frequency domain through the DWT module 204, and the subband image of the low frequency domain segment mapped in the frequency domain is split through the SPLIT module 205 to obtain four subband images Z of the second frequency domain segment (i.e., multiple sub-frequency domain segments of the low frequency domain segment). 11 , Z 12 , Z 13 and Z 14 . Further, the sub-band images Z of the four second frequency domain segments are 11 , Z 12 , Z 13 and Z 14 The corresponding denoising networks (FUNET networks) 2031, 2032, 2033 and 2034 are used for denoising respectively to obtain the intermediate denoised sub-band images Z' corresponding to the four second frequency domain segments. 11 , Z' 12 , Z' 13 and Z' 14 Further, based on the intermediate denoised sub-band images Z' corresponding to the four second frequency domain segments 11 , Z' 12 , Z' 13 and Z' 14 , and the initial denoised sub-band image corresponding to the low-frequency segment, thereby obtaining the target denoised sub-band image Z"1 corresponding to the low-frequency segment.

[0092] In some embodiments, in step B2, based on the intermediate denoised subband images of at least two second frequency domain segments and the initial denoised subband image corresponding to the first frequency domain segment where N is equal to 1, a target denoised subband image corresponding to the first frequency domain segment where N is equal to 1 is obtained, which can be achieved by the following steps.

[0093] Step B21, performing splicing and fusion processing and cross-band fusion processing on the intermediate denoised sub-band images of at least two second frequency domain segments to obtain a cross-band denoised sub-band fusion image corresponding to the first frequency domain segment where N is equal to 1;

[0094] Step B22: perform addition fusion processing on the cross-band denoised sub-band fusion image corresponding to the first frequency domain segment where N is equal to 1 and the initial denoised sub-band image to obtain a target denoised sub-band fusion image corresponding to the first frequency domain segment where N is equal to 1.

[0095] In an embodiment of the present application, the stitching and fusion processing can be implemented by a stitching (CONCAT) module. When different denoising networks extract features of different aspects or levels of the image that needs to be denoised in different frequency domain segments, the denoised image obtained can be stitched along the specified dimension through the stitching module to provide richer feature information for subsequent processing. It should be noted that the stitching module retains the original image information of all input denoised images of different frequency domain segments, and only merges them in space. In subsequent processing, the original image information of different frequency domain segments can be aggregated across frequency bands.

[0096] In the embodiment of the present application, the cross-band fusion process is the process of re-fusing sub-band images of different frequency domain segments. Since the sub-band images of different frequency domain segments respectively contain the information of the image in different frequency ranges (the low-frequency part contains the general outline and smooth area information of the image, and the high-frequency part contains the edge, texture and other detail information of the image), through cross-band fusion, the information in these different frequency ranges can be integrated to obtain a result containing more comprehensive and rich image information.

[0097] In the embodiment of the present application, the cross-band fusion processing can be implemented by a cross-band fusion network, which can be a traditional fully convolutional neural network, such as a UNET network, including multiple convolutions, downsampling and upsampling. The cross-band fusion processing can also be a network with different depths of cross-band information aggregation modules, such as a UNET network structure with frequency domain information (Frequency UNET, FUNET).

[0098] In an embodiment of the present application, the additive fusion processing can be implemented by an addition (ADD) module. After obtaining the cross-band denoised sub-band fusion image corresponding to the first frequency domain segment where N is equal to 1, the cross-band denoised sub-band fusion image corresponding to the first frequency domain segment where N is equal to 1 and the elements at corresponding positions in the initial denoised sub-band image corresponding to the first frequency domain segment where N is equal to 1 are added to achieve deep fusion of information of the two images of the first frequency domain segment where N is equal to 1, so that each element in the output image contains information from the elements at corresponding positions of the two images.

[0099] In the embodiment of the present application, after obtaining the intermediate denoised subband images of at least two second frequency domain segments, the electronic device performs a splicing and fusion process on the intermediate denoised subband images of at least two second frequency domain segments to obtain a first intermediate denoised subband fused image; further, the first intermediate denoised subband fused image is subjected to a cross-band fusion process to obtain a cross-band denoised subband fused image corresponding to the first frequency domain segment where N is equal to 1. Furthermore, the cross-band denoised subband fused image corresponding to the first frequency domain segment where N is equal to 1 is added to the elements at corresponding positions in the initial denoised subband image corresponding to the first frequency domain segment where N is equal to 1 to obtain a target denoised subband fused image corresponding to the first frequency domain segment where N is equal to 1.

[0100] Step B23: Perform spatial domain mapping on the target denoised sub-band fusion image corresponding to the first frequency domain segment where N is equal to 1, to obtain a target denoised sub-band image corresponding to the first frequency domain segment where N is equal to 1.

[0101] In the embodiment of the present application, the frequency domain image is mapped to the spatial domain by inverse wavelet transform, such as inverse discrete wavelet transform (IWT) or inverse Haar wavelet transform (IHWT). Of course, the frequency domain image can also be mapped to the spatial domain by inverse Fourier transform. This application does not impose any specific restrictions. It should be noted that the spatial domain mapping of the frequency domain image is the inverse of the frequency domain mapping of the spatial domain image. If the aforementioned frequency domain mapping adopts discrete wavelet transform, the spatial domain mapping adopts discrete wavelet transform; if the aforementioned frequency domain mapping adopts Haar wavelet transform, the spatial domain mapping adopts inverse Haar wavelet transform.

[0102] In a feasible scenario, continue to refer to Figure 2 , and then obtain the intermediate denoised subband images Z' corresponding to the four second frequency domain segments 11 , Z' 12 , Z' 13 and Z' 14 Then, the intermediate denoised sub-band images Z' corresponding to the four second frequency domain segments are 11, Z' 12 , Z' 13 and Z' 14 The CONCAT module 207 is used to perform splicing along the specified dimension, and the image output by the CONCAT module 207 is subjected to cross-channel information aggregation through the FUNET neural network 208 to obtain a cross-band denoised sub-band fusion image of the low-frequency domain segment; further, the ADD module 209 is used to add the pixel values ​​at corresponding positions of the cross-band denoised sub-band fusion image of the low-frequency domain segment and the initial denoised sub-band image of the low-frequency domain segment to obtain a target denoised sub-band fusion image corresponding to the low-frequency domain segment; finally, the target denoised sub-band fusion image corresponding to the low-frequency domain segment is spatially mapped through the IWT module 210 to obtain a target denoised sub-band image Z"1 corresponding to the low-frequency domain segment. In this way, a deep exploration and cross-band information aggregation are performed on the initial denoised sub-band image of the low-frequency domain segment to achieve a good denoising effect.

[0103] It should be noted that the above-mentioned low-frequency subband image to be denoised undergoes two frequency domain mapping and denoising processes. After each DWT frequency domain mapping, the length and width of the image will be reduced to 1 / 2 of the original after each frequency domain mapping. Therefore, in the denoising network corresponding to the two frequency domain mappings, for the two denoising processes, the time domain and frequency domain information in the image can be extracted from different scales. For example, some more global features or information are not significant when the image size is relatively large, but will be more significant when the image size is smaller.

[0104] In one scenario, among the multiple FUNET networks in the first denoising network, at least one FUNET network can be an Identity network whose input is equal to the output. Figure 2 For example, when the FUNET network 2031 in the first denoising network can be an Identity network with an input equal to an output, other FUNET networks can be ordinary neural networks. Since the FUNET network 2031 is an Identity network with an input equal to an output, the sub-band image to be denoised in the low-frequency domain segment is equivalent to undergoing two consecutive frequency domain mappings and one denoising process. In the denoising network corresponding to the second frequency domain mapping (such as FUNET networks 2061-2064), the time domain and frequency domain information in the image are extracted from different scales. It should be noted that whether the FUNET network 2031 is an Identity network or an ordinary neural network needs to be determined based on actual effects and other constraints.

[0105] Step B3: If N is greater than 1, based on the initial denoised subband image corresponding to the first frequency domain segment where N is greater than 1, perform N times of frequency domain mapping, denoising, fusion processing and spatial domain mapping to obtain a target denoised subband image corresponding to the first frequency domain segment where N is greater than 1.

[0106] In some embodiments, in step B3, if N is greater than 1, based on the initial denoised subband image corresponding to the first frequency domain segment where N is greater than 1, frequency domain mapping, denoising processing, fusion processing and spatial domain mapping are performed N times to obtain the target denoised subband image corresponding to the first frequency domain segment where N is greater than 1, which can be done through the following steps.

[0107] Step B31, performing an nth frequency domain mapping and denoising process on the initial denoised subband image corresponding to the first frequency domain segment where N is greater than 1, to obtain an nth intermediate denoised subband image of at least two nth second frequency domain segments; wherein n=1;

[0108] Step B32: performing an nth frequency domain mapping and denoising process on at least one n-1th intermediate denoised subband image of the n-1th second frequency domain segment to obtain at least two nth intermediate denoised subband images of the nth second frequency domain segment; wherein the value of n is any positive integer from 2 to N;

[0109] Step B33: after the Nth frequency domain mapping and denoising process, an Nth intermediate denoised subband image of at least two Nth second frequency domain segments is obtained; based on the at least two nth intermediate denoised subband images of the nth second frequency domain segments, an (N-n+1)th fusion process and spatial domain mapping are performed to obtain an n-1th target denoised subband image corresponding to the n-1th second frequency domain segment; wherein the value of n is all positive integers from N to 2;

[0110] Step B34: Based on the nth intermediate denoised subband images of at least two nth second frequency domain segments, perform the Nth feature fusion and spatial domain mapping to obtain the target denoised subband image corresponding to the first frequency domain segment where N is greater than 1, where n=1.

[0111] It can be understood that if N is greater than 1, first, the initial denoised subband image corresponding to the first frequency domain segment where N is greater than 1 is subjected to the nth (n=1)th frequency domain mapping and denoising process to obtain at least two nth (n=1)th intermediate denoised subband images of the nth (n=1)th second frequency domain segment. Further, the nth frequency domain mapping and denoising process is performed on at least one n-1th second frequency domain segment to obtain at least two nth intermediate denoised subband images of the nth second frequency domain segment. After the Nth frequency domain mapping and denoising process, the Nth intermediate denoised subband images of at least two Nth second frequency domain segments are obtained, and based on the nth intermediate denoised subband images of at least two nth second frequency domain segments, the (N-n+1)th fusion process and spatial domain mapping are performed to obtain the n-1th target denoised subband image corresponding to the n-1th second frequency domain segment; wherein the value of n is all positive integers from N to 2; finally, based on the nth intermediate denoised subband images of at least two nth second frequency domain segments, the Nth feature fusion and spatial domain mapping are performed to obtain the target denoised subband image corresponding to the first frequency domain segment where N is greater than 1, wherein n = 1. In this way, multiple depth explorations and cross-band information aggregation are performed on the initial denoised subband image of at least one first frequency domain segment to achieve a better denoising effect.

[0112] In some embodiments, step 103 performs fusion processing and spatial domain mapping based on at least two target denoised subband images of the first frequency domain segments to obtain a target denoised image, which can be achieved through the following steps.

[0113] The target denoised subband images based on at least two first frequency domain segments are stitched and fused to obtain a target intermediate denoised subband fused image; the target intermediate denoised subband fused image is cross-band fused to obtain a target cross-band denoised subband fused image; the target cross-band denoised subband fused image is spatially mapped to obtain a target denoised image.

[0114] In the embodiment of the present application, the splicing fusion processing can be implemented by a splicing (CONCAT) module, and the cross-band fusion processing can be implemented by a cross-band fusion network. The cross-band fusion network can be a traditional full convolutional neural network, such as a UNET network, which includes multiple convolutions, downsampling and upsampling. The cross-band fusion processing can also be a cross-band fusion network with different depths of a cross-band information aggregation module, such as a UNET network structure with frequency domain information.

[0115] In the embodiment of the present application, the description of the spatial domain mapping is similar to the above and will not be repeated here.

[0116] In the embodiment of the present application, after obtaining the target denoised sub-band images of at least two first frequency domain segments, the electronic device performs splicing and fusion processing on the target denoised sub-band images of at least two first frequency domain segments to obtain a target intermediate denoised sub-band fusion image; further, the target intermediate denoised sub-band fusion image is subjected to cross-band fusion processing to obtain a target cross-band denoised sub-band fusion image. Finally, the target cross-band denoised sub-band fusion image is subjected to spatial domain mapping to obtain a target denoised image of the image to be denoised.

[0117] In a feasible scenario, continue to refer to Figure 2 After obtaining the target denoising subband image Z"1 corresponding to the low-frequency segment, the target denoising subband image Z"2 corresponding to the horizontal high-frequency segment, the target denoising subband image Z"3 corresponding to the vertical high-frequency segment, and the target denoising subband image Z"4 corresponding to the diagonal high-frequency segment, the four target denoising subband images Z" 11 , Z” 12 , Z” 13 and Z" 14 The CONCAT module 211 is used to perform splicing along the specified dimension to obtain the target intermediate denoised sub-band fusion image; further, the target intermediate denoised sub-band fusion image is subjected to cross-channel information aggregation through the FUNET neural network 212 to obtain the target cross-band denoised sub-band fusion image; finally, the target cross-band denoised sub-band fusion image is subjected to spatial domain mapping through the IWT module 213 to obtain the target denoised image. In this way, the frequency domain characteristics of the input image to be denoised are fully utilized, with both in-depth exploration of sub-bands of different frequency bands and information aggregation across frequency bands to achieve a good denoising effect.

[0118] In some embodiments, reference Figure 5 As shown, the image noise reduction method may further include the following steps:

[0119] Step 501: Obtain adjustment parameters of a reference frequency domain segment.

[0120] The reference frequency domain segment includes one or more of at least two first frequency domain segments.

[0121] In an embodiment of the present application, the adjustment parameter is used to adjust the image intensity, such as scaling the image intensity, and the adjustment parameter may be a pre-input parameter. It should be noted that the image intensity may be the intensity of the content of each frequency band in the image. Exemplarily, if it is determined that the denoised image retains as many high-frequency details as possible, the adjustment parameter may be set to be greater than or equal to 1 for the three sub-bands of high-frequency information corresponding to the horizontal high-frequency domain segment, the vertical high-frequency domain segment, and the diagonal high-frequency domain segment, so as to achieve that the image output by the network has more high-frequency details.

[0122] In an embodiment of the present application, the reference frequency domain segment may be one or more of at least two first frequency domain segments, and the reference frequency domain segment may also be a sub-frequency domain segment of different levels obtained by dividing the first frequency domain segment once or multiple times. Exemplarily, the reference frequency domain segment may be one or more of the sub-frequency domain segments (corresponding to the above-mentioned second frequency domain segment) after dividing at least one first frequency domain segment once, and the reference frequency domain segment may be one or more of the sub-frequency domain segments obtained by dividing at least one first frequency domain segment twice or multiple times. This application does not impose specific restrictions on the reference frequency domain segment.

[0123] Step 502: Perform denoising processing based on the adjustment parameters of the reference frequency domain segment and the reference sub-band image to be denoised of the reference frequency domain segment to obtain a reference denoised sub-band image of the reference frequency domain segment.

[0124] The reference denoised sub-band image includes the initial denoised sub-band image.

[0125] In the embodiment of the present application, based on the adjustment parameters of the reference frequency domain segment, the denoising process can be implemented by a second denoising network, the second denoising network includes a first denoising network, and the first denoising network is used to perform denoising on the sub-band images of each frequency domain segment. Exemplarily, the network structure of the second denoising network is as follows: Figure 6 As shown, the second noise reduction network includes the first noise reduction network, a multiplication module and an addition module.

[0126] In an embodiment of the present application, the reference sub-band image to be denoised may be an image that needs to be denoised by a second denoising network corresponding to the reference frequency domain segment, and the reference denoised sub-band image may be an image obtained after the reference sub-band image to be denoised is denoised by the second denoising network.

[0127] It can be understood that step 502 performs noise reduction processing based on the adjustment parameters of the reference frequency domain segment and the reference sub-band image to be denoised of the reference frequency domain segment to obtain the reference denoised sub-band image of the reference frequency domain segment, which can be implemented through the following steps.

[0128] Step D1, performing initial denoising processing on the reference denoised sub-band image of the reference frequency domain segment to obtain an initial reference denoised sub-band image of the reference frequency domain segment;

[0129] Step D2: adjusting the initial reference denoised sub-band image based on the adjustment parameter to obtain an adjusted initial reference denoised sub-band image;

[0130] Step D3: fusing the adjusted initial reference denoised sub-band image and the initial reference denoised sub-band image of the reference frequency domain segment to obtain a reference denoised sub-band image of the reference frequency domain segment.

[0131] In an embodiment of the present application, the initial denoising process can be implemented by the first denoising network in the second denoising network, and the reference denoising subband image in the reference frequency domain segment is subjected to initial denoising process by the first denoising network in the second denoising network to obtain an initial reference denoising subband image; further, the adjustment parameter and the initial reference denoising subband image are fused by the multiplication module in the second denoising network to obtain an adjusted initial reference denoising subband image; finally, the adjusted initial reference denoising subband image and the initial reference denoising subband image of the reference frequency domain segment are fused by the addition module in the second denoising network to obtain a reference denoising subband image of the reference frequency domain segment. In this way, the embodiment of the present application can add an adjustable adjustment parameter to each frequency domain segment to achieve precise adjustment of the final denoising effect in the frequency domain.

[0132] Reference Figure 7 , Figure 7 An image noise reduction method provided in an embodiment of the present application can be applied to any electronic device with image noise reduction capability. The method comprises the following steps:

[0133] Step 701: Obtain adjustment parameters of the image to be denoised and the reference frequency domain segment.

[0134] The reference frequency domain segment is one or more of at least two first frequency domain segments, and the at least two first frequency domain segments are frequency domain segments obtained by dividing the frequency domain segment where the image to be denoised is located after frequency domain mapping.

[0135] Step 702: Based on the adjustment parameters of the reference frequency domain segment, frequency domain mapping and noise reduction processing are performed on the image to be denoised to obtain at least two initial denoised sub-band images of the first frequency domain segment.

[0136] Step 703: Determine the target denoised sub-band images corresponding to the initial denoised sub-band images, and obtain at least two target denoised sub-band images in the first frequency domain segment.

[0137] The target denoised sub-band image is obtained by performing N frequency domain mapping and denoising processes on the initial denoised sub-band image, where N≥0 and N is an integer.

[0138] In the embodiment of the present application, the number threshold can be 0.

[0139] In some embodiments, N corresponding to different first frequency domain segments are the same or different, and N corresponding to at least one first frequency domain segment is greater than or equal to a number threshold.

[0140] In the embodiment of the present application, the number of times the initial denoised subband images of different first frequency domain segments are re-frequency-domain mapped and denoised is the same or different. Exemplarily, if the same, the number of frequency-domain mapping times N of the images of all first frequency domain segments can be any one of 0, 1, 2, 3, etc., and the value of N can be preset. Of course, N can also be set to other numbers other than the above examples. If different, the number of frequency-domain mapping times N of the images of some first frequency domain segments can be 0, and the number of frequency-domain mapping times N of the images of the remaining first frequency domain segments can be any number greater than 0. It should be noted that the above examples are only illustrative, and the embodiments of the present application can also be set according to other situations, as long as the N corresponding to different first frequency domain segments is the same or different.

[0141] Step 704: Perform fusion processing and spatial domain mapping based on at least two target denoised subband images of the first frequency domain segments to obtain a target denoised image.

[0142] In one possible scenario, refer to Figure 2 and Figure 8 As shown, a diagonal high-frequency scaling factor (i.e., adjustment parameter) is added as the second input, which is input into the second denoising network 2035 corresponding to the diagonal high-frequency domain segment, and an initial denoised subband image Z'4 corresponding to the diagonal high-frequency domain segment is obtained. Further, when the initial denoised subband image Z'1 of the low-frequency domain segment, the initial denoised subband image Z'2 of the horizontal high-frequency domain segment, and the initial denoised subband image Z'3 of the vertical high-frequency domain segment are obtained, Figure 8 It can be seen that no further frequency domain mapping and noise reduction processing is performed on each frequency domain segment. At this time, the initial noise reduction subband image Z'1 of the low frequency domain segment can be used as the target noise reduction subband image of the low frequency domain segment, the initial noise reduction subband image Z'2 of the horizontal high frequency domain segment can be used as the target noise reduction subband image of the horizontal high frequency domain segment, the initial noise reduction subband image Z'3 of the vertical high frequency domain segment can be used as the target noise reduction subband image of the vertical high frequency domain segment, and the initial noise reduction subband image Z'4 of the diagonal high frequency domain segment can be used as the target noise reduction subband image of the diagonal high frequency domain segment. Then, based on the target noise reduction subband images of each frequency domain segment, a variety of different fusion processes and one spatial domain mapping are performed to obtain the target noise reduction image. In this way, an adjustable scaling parameter is added to each frequency band to achieve precise and adjustable final noise reduction effect in the frequency domain.

[0143] It should be noted that, for the description of the same steps and the same contents in this embodiment as those in other embodiments, reference can be made to the description in other embodiments and will not be repeated here.

[0144] In some embodiments, the denoising networks involved above (including the first denoising network and the second denoising network) are pre-trained with sample data, and the sample data includes: a target true value image corresponding to the true value image, and a noisy image obtained by superimposing noise on the true value image, wherein the target true value image is an image obtained after edge enhancement is performed on the true value image at least based on different edge enhancement strategies.

[0145] In the embodiments of the present application, the true value image can be understood as an original image that is not affected by noise or other interference and can truly reflect the actual situation of the shooting scene or object. The true value image is an image in an ideal state, which is often used as a reference standard in image processing to evaluate the processing effects of various algorithms such as noise reduction algorithms on images.

[0146] In the embodiment of the present application, since the noise image is obtained by superimposing noise on the true value image, the true value image can be regarded as the image obtained after denoising the noise image, and the true value image after different edge enhancements is fused to obtain the target true value image, which is used as part of the sample data to train the denoising network, and the network can learn more feature information about the edge of the image. When the noise image is actually denoised, the edge details of the image can be better restored, making the denoised image clearer, avoiding the edge blur during the denoising process, thereby improving the visual quality of the image. At the same time, the edge of the image contains important structural information of the object. Through training, the denoising network learns the edge features under different edge enhancement strategies, which helps to accurately identify and retain these structural information during the denoising process, which is of great significance for image understanding and analysis tasks. In addition, through the training of a large number of different noise images and corresponding target true value images, the denoising network can better understand the relationship between noise and real image information, so that when encountering a noise image that has not appeared in the training set, it can also accurately remove the noise and restore a near-real image, enhancing the generalization ability of the network in different scenarios.

[0147] In some embodiments, the process of obtaining sample data is combined with Fig. 9 Provide explanation.

[0148] Step 901: Acquire initial sample data, where the initial sample data includes a true value image and a noise image;

[0149] Step 902: adopt at least two edge enhancement strategies to perform edge enhancement on the true value image respectively, and obtain at least two enhanced true value images corresponding to the true value image.

[0150] In the embodiments of the present application, the edge enhancement strategy can be understood as a strategy of an image processing method for highlighting the edge of an object in an image. The edge of an image refers to an area in an image where the grayscale value changes dramatically, representing the boundary of an object or the dividing line between different areas. The edge enhancement strategy processes the image through a specific algorithm to increase the contrast between the edge and non-edge areas, making the edge more obvious.

[0151] In practical applications, different edge enhancement strategies may be strategies for enhancing an image based on different parameters in the same algorithm, or strategies for enhancing an image using different algorithms, or a combination of the two.

[0152] Common edge enhancement strategies include: Sobel operator, Prewitt operator, Laplacian operator and Laplacian of Gaussian (LoG) operator, etc.

[0153] The Sobel operator and Prewitt operator detect edges by calculating the gradient value of each pixel in the image. A large gradient value indicates a dramatic change in grayscale, which means that an edge may exist. For example, the Sobel operator performs convolution operations in the horizontal and vertical directions respectively to obtain the horizontal and vertical gradient components, and then calculates the total gradient amplitude and direction to highlight the edge information.

[0154] The above-mentioned Laplace operator is a second-order derivative operator that is very sensitive to edges and noise in images. The Laplace operator generates zero crossings at the edges by calculating the second-order derivative of the image, thereby detecting the edges. Because it is sensitive to noise, the image usually needs to be smoothed before use.

[0155] The above LoG operator needs to use a Gaussian function to smooth the image first, and then apply the Laplace operator to detect the edge. By combining the smoothing characteristics of the Gaussian filter and the edge detection ability of the Laplace operator, the influence of noise on edge detection can be reduced to a certain extent, and more accurate edge information can be obtained.

[0156] In the embodiment of the present application, the enhanced true value image is an image obtained by applying at least one edge enhancement strategy to the true value image. Since different edge enhancement strategies emphasize different edge features and enhancement effects, at least two different enhanced true value images will be obtained by processing the true value image using at least two edge enhancement strategies. The enhanced true value image highlights the edge features in different aspects or degrees on the basis of retaining the basic information of the true value image, and can be used for subsequent image processing tasks, such as training a denoising network as sample data to help the network learn rich image edge features, thereby better restoring and retaining the edge details of the image during the denoising process.

[0157] Step 903: Perform frequency domain mapping and fusion processing based on at least two enhanced true value images to obtain a target true value image.

[0158] In the embodiment of the present application, the target true value image is an image obtained after frequency domain mapping and fusion processing based on at least two enhanced true value images. Since the target true value image is obtained based on multiple enhanced true value images, compared with a single enhanced true value image, the target true value image contains more comprehensive and accurate edge and detail features, and can be used as sample data for training the denoising network, helping the network to learn richer image features, so as to improve the denoising and detail recovery capabilities of noisy images.

[0159] In some embodiments, step 903 performs frequency domain mapping and fusion processing based on at least two enhanced true value images to obtain a target true value image, which can be achieved through the following steps.

[0160] Step E1: Perform frequency domain mapping on at least two enhanced true value images to obtain enhanced true value sub-band images of at least two frequency domain segments under the same enhanced true value image.

[0161] In the embodiment of the present application, the enhanced true value sub-band image can be understood as an image of different frequency domain segments obtained by dividing the enhanced true value image according to the frequency range after the enhanced true value image is frequency-domain mapped. It should be noted that the enhanced true value sub-band image contains the information of the original enhanced true value image in different frequency ranges, which helps to analyze and process the image more carefully.

[0162] Step E2: Perform weighted fusion processing on at least two enhanced true value sub-band images corresponding to the same frequency domain segment to obtain fused enhanced true value sub-band images of at least two frequency domain segments.

[0163] In an embodiment of the present application, for at least two frequency domain segments, at least two enhanced true value sub-band images corresponding to the same frequency domain segment are weighted fused to obtain a fused enhanced true value sub-band image corresponding to the frequency domain segment, thereby obtaining a fused enhanced true value sub-band image of at least two frequency domain segments.

[0164] It should be noted that a weight can be assigned to each enhanced true subband image corresponding to the same frequency domain segment. This weight reflects the importance of the subband image in the fusion process. Then, the enhanced true subband images are fused according to their corresponding weights to obtain the fused enhanced true subband images. Through weighted fusion, the advantages of multiple enhanced true subband images can be combined to highlight more important information in the image, such as edges and textures, so that the target true image retains rich information while the key features are more obvious, which helps to improve the visual quality and recognizability of the image, and is of great significance for image analysis and understanding tasks.

[0165] In some embodiments, a weight is assigned to each enhanced true value sub-band image corresponding to the same frequency domain segment, which can be preset according to task requirements or determined based on image information in each enhanced true value sub-band image.

[0166] Step E3: Based on the fused enhanced true value sub-band images of at least two frequency domain segments, cross-band fusion processing and spatial domain mapping are performed to obtain a target true value image.

[0167] In an embodiment of the present application, since the enhanced true value sub-band images after fusion of different frequency domain segments respectively contain the information of the image within different frequency ranges, the image information of the enhanced true value sub-band images after fusion of at least two frequency domain segments can be integrated through cross-band fusion to obtain a cross-frequency domain enhanced true value sub-band image containing more comprehensive and rich image information; further, the cross-frequency domain enhanced true value sub-band image is converted from the frequency domain back to the spatial domain, so that the processed target true value image can be intuitively observed and used.

[0168] From the above, it can be seen that the target true value image of the embodiment of the present application combines the advantageous information of multiple enhanced true value images, and contains rich edge and detail features. When it is used as sample data for training the denoising network, it can provide the network with more comprehensive and accurate image feature information, helping the network to better learn the difference between noise and real images, thereby improving the network's denoising ability and the ability to restore image details. Compared with using a single enhanced true value image or other ordinary sample data, using the target true value image for training can make the denoising network more robust and accurate when facing various complex noises, improving the overall performance and generalization ability of the network.

[0169] Step 904: Use the target true value image and the noise image as sample data.

[0170] It should be noted that, for the description of the same steps and the same contents in this embodiment as those in other embodiments, reference can be made to the description in other embodiments and will not be repeated here.

[0171] It should be noted that, for the description of the same steps and the same contents in this embodiment as those in other embodiments, reference can be made to the description in other embodiments and will not be repeated here.

[0172] This application takes an implementation scenario as an example. Fig.10 As shown, an image noise reduction method provided in an embodiment of the present application is further described. When applied to an electronic device, the method can be implemented in the following manner.

[0173] Step 1001: perform frequency domain mapping on the input noisy image to obtain sub-band information of multiple different frequency bands, and the sub-band information of multiple different frequency bands respectively passes through the corresponding denoising network including frequency domain feature extraction to obtain initial denoised sub-band information of multiple different frequency bands.

[0174] Here, the sub-band information (corresponding to the sub-band images to be denoised) of multiple different frequency bands (corresponding to different frequency domain segments or different frequency characteristics) includes low-frequency sub-band information Z1 (corresponding to the sub-band image to be denoised in the low-frequency domain segment), horizontal high-frequency sub-band information Z2 (corresponding to the sub-band image to be denoised in the horizontal high-frequency domain segment), vertical high-frequency sub-band information Z3 (corresponding to the sub-band image to be denoised in the vertical high-frequency domain segment) and diagonal high-frequency sub-band information Z4 (corresponding to the sub-band image to be denoised in the diagonal high-frequency domain segment).

[0175] Here, the input noisy image (corresponding to the image to be denoised) is first mapped in the frequency domain through the DWT module, and the sub-band information of the frequency domain mapping is split into multiple sub-band information Z1, Z2, Z3 and Z4 of different frequency bands through the SPLIT module. Further, the sub-band information Z1, Z2, Z3 and Z4 of multiple different frequency bands are respectively passed through the corresponding denoising network including frequency domain feature extraction to obtain multiple initial denoised sub-band information (corresponding to the above-mentioned initial denoised sub-band images) Z'1, Z'2, Z'3 and Z'4 of different frequency bands.

[0176] Step 1002: further perform frequency domain mapping on the initial denoised subband information of the low frequency to obtain sub-subband information of multiple different sub-bands, and the sub-subband information of multiple different sub-bands respectively passes through the corresponding denoising network including frequency domain feature extraction to obtain intermediate denoised subband information of multiple different sub-bands.

[0177] Here, the low-frequency initial denoising sub-band information Z'1 is further frequency-domain mapped, and the sub-sub-band information Z of multiple different sub-bands obtained after mapping is converted into 11 , Z 12 , Z 13 and Z 14 , respectively, through the denoising network including frequency domain feature extraction, to obtain the processed sub-sub-band information Z' of multiple different sub-bands 11 , Z' 12 , Z' 13 and Z' 14 .

[0178] Step 1003: After concatenating and aggregating the processed sub-sub-band information of multiple different sub-frequency bands, the information is added to the initial noise reduction sub-band information of the low frequency band, and then spatially mapped to obtain the target noise reduction sub-band information of the low frequency band.

[0179] Here, the processed sub-subband information Z' of multiple different sub-bands is11 , Z' 12 , Z' 13 and Z' 14 After splicing and cross-band information aggregation, it is added to the initial denoising sub-band information Z'1 of the low-frequency band, and then spatially mapped through the IWT module to obtain the target denoising sub-band information Z"1 of the low-frequency band.

[0180] Step 1004: The initial denoising subband information of the horizontal high frequency, the initial denoising subband information of the vertical high frequency and the initial denoising subband information of the diagonal high frequency are respectively used as the respective target denoising subband information, and the target denoising subband information of the low frequency band, the target denoising subband information of the horizontal high frequency and the target denoising subband information of the vertical high frequency are subjected to splicing, cross-band information aggregation and spatial domain mapping to obtain the denoising result of the noisy image.

[0181] Here, the initial denoising subband information Z'2 of horizontal high frequency, the initial denoising subband information Z'3 of vertical high frequency and the initial denoising subband information Z'4 of diagonal high frequency are respectively used as the target denoising subband information, and the target noise subband information Z"1 of the low frequency band, the target denoising subband information Z"1 of the horizontal high frequency, the target denoising subband information Z"3 of the vertical high frequency and the target denoising subband information Z"4 of the diagonal high frequency are spliced, aggregated across frequency bands and mapped in the spatial domain of the IWT module to obtain the denoised image of the noisy image, that is, the denoising result.

[0182] It should be noted that the embodiment of the present application is aimed at the characteristics of the two-dimensional image after wavelet transformation. This solution further splits the subband representing the low-frequency component after the first discrete wavelet transform, applies the second discrete wavelet transform DWT, and each frequency band information is respectively passed through the network, cross-frequency band information aggregation, and the second discrete wavelet inverse transform method to obtain the low-frequency component of the first wavelet transform after processing. The above discrete wavelet transform and discrete inverse wavelet transform can be implemented using convolution and other schemes, so the above process can be completely implemented through a deep neural network.

[0183] From the above, it can be seen that the embodiment of the present application fully utilizes the frequency domain characteristics of the input image to be denoised, with both in-depth exploration of sub-bands in different frequency bands and information aggregation across frequency bands, making full use of frequency domain and spatial domain information, and can retain the high-frequency details of the image as much as possible while denoising, thereby achieving a good noise reduction effect.

[0184] It should be noted that, for the description of the same steps and the same contents in this embodiment as those in other embodiments, reference can be made to the description in other embodiments and will not be repeated here.

[0185] Reference Fig.11 , Fig.11 An image noise reduction device provided in an embodiment of the present application includes:

[0186] An acquisition module 1101 is used to obtain an image to be denoised;

[0187] A processing module 1102 is used to perform frequency domain mapping and noise reduction processing on the image to be denoised, so as to obtain at least two initial denoised sub-band images of the first frequency domain segment;

[0188] A determination module 1103 is used to determine a target denoised sub-band image corresponding to each initial denoised sub-band image, and obtain at least two target denoised sub-band images of the first frequency domain segment, wherein the target denoised sub-band image is obtained by performing N frequency domain mapping and denoising processes based on the initial denoised sub-band image, where N≥0, and N is an integer;

[0189] The processing module 1102 is further configured to perform fusion processing and spatial domain mapping based on at least two target denoised sub-band images of the first frequency domain segments to obtain a target denoised image.

[0190] An embodiment of the present application provides an electronic device, including a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the program, some or all of the steps in the above method are implemented.

[0191] The embodiment of the present application provides a storage medium, which stores one or more computer programs, and the one or more computer programs can be executed by one or more processors to implement some or all of the steps in the above method. The storage medium can be transient or non-transient.

[0192] An embodiment of the present application provides a computer program, including a computer-readable code. When the computer-readable code is run in an electronic device, a processor in the electronic device executes some or all of the steps for implementing the above method.

[0193] The embodiment of the present application provides a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program, and when the computer program is read and executed by a computer, some or all of the steps in the above method are implemented. The computer program product can be implemented specifically by hardware, software or a combination thereof. In some embodiments, the computer program product is specifically embodied as a computer storage medium, and in other embodiments, the computer program product is specifically embodied as a software product, such as a software development kit (SDK) and the like.

[0194] It should be noted here that the description of the various embodiments above tends to emphasize the differences between the various embodiments, and the same or similar aspects can be referenced to each other. The description of the above device, storage medium, computer program and computer program product embodiments is similar to the description of the above method embodiment, and has similar beneficial effects as the method embodiment. For technical details not disclosed in the embodiments of the device, storage medium, computer program and computer program product of this application, please refer to the description of the method embodiment of this application for understanding.

[0195] Fig.12 A hardware entity diagram of an electronic device provided in an embodiment of the present application, such as Fig.12 As shown, the hardware entity of the electronic device 12 includes: a processor 1201 and a memory 1202, wherein the memory 1202 stores a computer program that can be run on the processor 1201, and the processor 1201 implements the following steps when executing the program:

[0196] Obtaining an image to be denoised, performing frequency domain mapping and denoising processing on the image to be denoised, and obtaining at least two initial denoised sub-band images of a first frequency domain segment;

[0197] Determine a target denoised subband image corresponding to each initial denoised subband image, and obtain at least two target denoised subband images of the first frequency domain segment, wherein the target denoised subband image is obtained by performing frequency domain mapping and denoising processing N times based on the initial denoised subband image, where N≥0, and N is an integer;

[0198] Based on at least two target denoised subband images of the first frequency domain segment, fusion processing and spatial domain mapping are performed to obtain a target denoised image.

[0199] Among them, the memory 1202 stores a computer program that can be run on the processor. The memory 1202 is configured to store instructions and applications executable by the processor 1201. It can also cache data to be processed or processed by the processor 1201 and various modules in the electronic device 12 (for example, image data, audio data, voice communication data, and video communication data), which can be implemented through flash memory (FLASH) or random access memory (Random Access Memory, RAM).

[0200] The processor 1201 generally controls the overall operation of the electronic device 12 .

[0201] An embodiment of the present application provides a computer storage medium, which stores one or more programs. The one or more programs can be executed by one or more processors to implement the steps of the image denoising method of any of the above embodiments.

[0202] It should be noted here that the description of the above storage medium and device embodiments is similar to the description of the above method embodiments, and has similar beneficial effects as the method embodiments. For technical details not disclosed in the storage medium and device embodiments of this application, please refer to the description of the method embodiments of this application for understanding.

[0203] The processor may be at least one of an Application Specific Integrated Circuit (ASIC), a Digital Signal Processor (DSP), a Digital Signal Processing Device (DSPD), a Programmable Logic Device (PLD), a Field Programmable Gate Array (FPGA), a Central Processing Unit (CPU), a controller, a microcontroller, and a microprocessor. It is understandable that the electronic device that implements the functions of the processor may also be other, and the embodiments of the present application are not specifically limited.

[0204] The above-mentioned computer storage medium / memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a magnetic random access memory (FRAM), a flash memory (Flash Memory), a magnetic surface memory, an optical disk, or a compact disc read-only memory (CD-ROM) and the like; it can also be various terminals including one or any combination of the above-mentioned memories, such as mobile phones, computers, tablet devices, personal digital assistants, etc.

[0205] It should be understood that "one embodiment" or "an embodiment" mentioned throughout the specification means that specific features, structures or characteristics related to the embodiment are included in at least one embodiment of the present application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. In addition, these specific features, structures or characteristics can be combined in one or more embodiments in any suitable manner. It should be understood that in various embodiments of the present application, the size of the serial number of each step / process mentioned above does not mean the order of execution, and the execution order of each step / process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present application. The serial numbers of the embodiments of the present application mentioned above are for description only and do not represent the advantages and disadvantages of the embodiments.

[0206] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0207] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0208] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units; they may be located in one place or distributed on multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0209] In addition, all functional units in the embodiments of the present application may be integrated into one processing unit, or each unit may be a separate unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0210] A person skilled in the art can understand that all or part of the steps of implementing the above method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiment; and the aforementioned storage medium includes: mobile storage devices, read-only memories (ROM), magnetic disks or optical disks, etc., various media that can store program codes.

[0211] Alternatively, if the above-mentioned integrated unit of the present application is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can essentially or in other words, the part that contributes to the relevant technology can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for enabling a vehicle-mounted terminal (which can be a personal computer, server, or network device, etc.) to execute all or part of the methods described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0212] The above is only an implementation method of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. An image denoising method, characterized in that: The method comprises: Obtaining an image to be denoised, performing frequency domain mapping and denoising processing on the image to be denoised, and obtaining at least two initial denoised sub-band images of a first frequency domain segment; Determine a target denoised subband image corresponding to each of the initial denoised subband images, and obtain at least two target denoised subband images of the first frequency domain segment, wherein the target denoised subband image is obtained by performing frequency domain mapping and denoising processing N times based on the initial denoised subband image, where N≥0, and N is an integer; Based on the target denoised sub-band images of the at least two first frequency domain segments, fusion processing and spatial domain mapping are performed to obtain a target denoised image.

2. The method according to claim 1, characterized in that: The Ns corresponding to different first frequency domain segments are the same or different, and the N corresponding to at least one first frequency domain segment is greater than a times threshold.

3. The method according to claim 1, wherein determining the target denoised sub-band image corresponding to each of the initial denoised sub-band images comprises: If N is equal to 0, the initial denoised sub-band image corresponding to the first frequency domain segment where N is equal to 0 is determined as the target denoised sub-band image corresponding to the first frequency domain segment where N is equal to 0; If N is equal to 1, based on the initial denoised subband image corresponding to the first frequency domain segment with N equal to 1, perform frequency domain mapping and denoising processing once to obtain at least two intermediate denoised subband images of the second frequency domain segments; based on the intermediate denoised subband images of the at least two second frequency domain segments and the initial denoised subband image corresponding to the first frequency domain segment with N equal to 1, obtain the target denoised subband image corresponding to the first frequency domain segment with N equal to 1, wherein the first frequency domain segment with N equal to 1 includes the at least two second frequency domain segments; If N is greater than 1, based on the initial denoised subband image corresponding to the first frequency domain segment where N is greater than 1, frequency domain mapping, denoising processing, fusion processing and spatial domain mapping are performed N times to obtain a target denoised subband image corresponding to the first frequency domain segment where N is greater than 1.

4. The method according to claim 3, wherein obtaining the target denoised subband image corresponding to the first frequency domain segment where N is equal to 1 based on the intermediate denoised subband images of the at least two second frequency domain segments and the initial denoised subband image corresponding to the first frequency domain segment where N is equal to 1 comprises: Performing a splicing and fusion process on the intermediate denoised sub-band images of the at least two second frequency domain segments and a cross-band fusion process to obtain a cross-band denoised sub-band fusion image corresponding to the first frequency domain segment where N is equal to 1; Performing an addition fusion process on the cross-band denoised sub-band fused image and the initial denoised sub-band image corresponding to the first frequency domain segment where N is equal to 1, to obtain a target denoised sub-band fused image corresponding to the first frequency domain segment where N is equal to 1; The target denoised sub-band fusion image corresponding to the first frequency domain segment where N is equal to 1 is spatially mapped to obtain the target denoised sub-band image corresponding to the first frequency domain segment where N is equal to 1.

5. The method according to claim 3, wherein if N is greater than 1, based on the initial denoised sub-band image corresponding to the first frequency domain segment where N is greater than 1, performing N times of frequency domain mapping, denoising processing, fusion processing and spatial domain mapping to obtain a target denoised sub-band image corresponding to the first frequency domain segment where N is greater than 1, comprises: Performing an nth frequency domain mapping and denoising process on the initial denoised subband image corresponding to the first frequency domain segment where N is greater than 1, to obtain at least two nth intermediate denoised subband images of the nth second frequency domain segment; wherein n=1; Performing an n-th frequency domain mapping and denoising process on at least one n-1-th intermediate denoised subband image of the second frequency domain segment, to obtain at least two n-th intermediate denoised subband images of the second frequency domain segment; wherein the value of n is any positive integer from 2 to N; After the Nth frequency domain mapping and denoising process, an Nth intermediate denoised subband image of at least two Nth second frequency domain segments is obtained, and based on the at least two nth intermediate denoised subband images of the nth second frequency domain segments, an (N-n+1)th fusion process and spatial domain mapping are performed to obtain an n-1th target denoised subband image corresponding to the n-1th second frequency domain segment; wherein the value of n is all positive integers from N to 2; Based on the nth intermediate denoised subband images of the at least two nth second frequency domain segments, an Nth feature fusion and spatial domain mapping are performed to obtain a target denoised subband image corresponding to the first frequency domain segment where N is greater than 1, where n=1.

6. The method according to any one of claims 1 to 5, characterized in that: The method further comprises: Obtaining an adjustment parameter of a reference frequency domain segment, wherein the reference frequency domain segment includes one or more of the at least two first frequency domain segments; Based on the adjustment parameters of the reference frequency domain segment and the reference denoised subband image of the reference frequency domain segment, denoising processing is performed to obtain a reference denoised subband image of the reference frequency domain segment, wherein the reference denoised subband image includes the initial denoised subband image.

7. The method according to claim 6, characterized in that The step of performing denoising processing based on the adjustment parameter of the reference frequency domain segment and the reference sub-band image to be denoised of the reference frequency domain segment to obtain the reference denoised sub-band image of the reference frequency domain segment includes: Performing initial denoising processing on the reference to-be-denoised sub-band image of the reference frequency domain segment to obtain an initial reference denoised sub-band image of the reference frequency domain segment; Based on the adjustment parameter, the initial reference denoised sub-band image is adjusted to obtain an adjusted initial reference denoised sub-band image; The adjusted initial reference denoised sub-band image of the reference frequency domain segment and the initial reference denoised sub-band image are fused to obtain a reference denoised sub-band image of the reference frequency domain segment.

8. The method according to any one of claims 1 to 6, characterized in that: The fusion processing includes one or more of splicing fusion processing, cross-band fusion processing and addition fusion processing. The noise reduction processing is performed by a noise reduction network, and different frequency domain segments correspond to different noise reduction networks; The cross-band fusion processing is performed by a cross-band fusion network, wherein the cross-band fusion network includes networks of different depths of cross-band information aggregation modules; The noise reduction network includes a convolutional neural network, or a neural network of different depths including the cross-band information aggregation module.

9. The method according to claim 8, characterized in that The denoising network is pre-trained with sample data, and the sample data includes: a target true value image corresponding to the true value image, and a noisy image obtained by superimposing noise on the true value image, wherein the target true value image is an image obtained after edge enhancement is performed on the true value image at least based on different edge enhancement strategies.

10. The method according to claim 9, characterized in that The method further comprises: Acquire initial sample data, wherein the initial sample data includes the true value image and the noise image; Adopting at least two edge enhancement strategies to respectively perform edge enhancement on the true value image, so as to obtain at least two enhanced true value images corresponding to the true value image; Based on the at least two enhanced true value images, frequency domain mapping and fusion processing are performed to obtain the target true value image; The target true value image and the noise image are used as the sample data.

11. The method according to claim 10, characterized in that The performing frequency domain mapping and fusion processing based on the at least two enhanced true value images to obtain the target true value image includes: Performing frequency domain mapping on the at least two enhanced true value images to obtain enhanced true value sub-band images of at least two frequency domain segments under the same enhanced true value image; Performing weighted fusion processing on at least two enhanced true value sub-band images corresponding to the same frequency domain segment to obtain a fused enhanced true value sub-band image of the at least two frequency domain segments; Based on the enhanced true value sub-band image after fusion of the at least two frequency domain segments, cross-frequency band fusion processing and spatial domain mapping are performed to obtain the target true value image.

12. The method according to any one of claims 1 to 6, characterized in that: The performing frequency domain mapping and noise reduction processing on the image to be denoised to obtain at least two initial denoised sub-band images of the first frequency domain segment includes: Performing frequency domain mapping on the image to be denoised to obtain at least two sub-band images to be denoised in the first frequency domain segment; The sub-band images to be denoised of the at least two first frequency domain segments are subjected to denoising processing respectively to obtain initial denoised sub-band images of the at least two first frequency domain segments.

13. The method according to any one of claims 1 to 6, characterized in that: The step of performing fusion processing and spatial domain mapping on the target denoised subband images based on the at least two first frequency domain segments to obtain a target denoised image includes: Performing a splicing and fusion process on the target denoised sub-band images of the at least two first frequency domain segments to obtain a target intermediate denoised sub-band fused image; Performing cross-band fusion processing on the target intermediate denoised sub-band fused image to obtain a target cross-band denoised sub-band fused image; The target cross-band denoised sub-band fusion image is spatially mapped to obtain the target denoised image.

14. An image noise reduction device, characterized in that: The device comprises: An acquisition module, used for acquiring an image to be denoised; A processing module, configured to perform frequency domain mapping and noise reduction processing on the image to be denoised, to obtain at least two initial denoised sub-band images of the first frequency domain segment; A determination module, configured to determine a target denoised subband image corresponding to each of the initial denoised subband images, and obtain at least two target denoised subband images of the first frequency domain segment, wherein the target denoised subband image is obtained by performing N frequency domain mapping and denoising processing on the initial denoised subband image, where N≥0, and N is an integer; The processing module is further used to perform fusion processing and spatial domain mapping based on the target denoised sub-band images of the at least two first frequency domain segments to obtain a target denoised image.

15. An electronic device, characterized in that: The electronic device comprises: a memory and a processor; The memory stores a computer program executable on the processor. When the processor executes the computer program, the image noise reduction method according to any one of claims 1 to 13 is implemented.

16. A storage medium, characterized in that: The storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the image noise reduction method according to any one of claims 1 to 13.

Images