Image processing methods, apparatus, electronic devices and storage media

Abstract

This disclosure proposes an image processing method, apparatus, electronic device, and storage medium, belonging to the field of image processing technology. The method, in response to acquiring an original photoelectric signal image, performs a frequency division operation on the original photoelectric signal image to obtain a texture noise component and a main structure component. By physically separating the texture noise and the main structure, feature extraction is performed in parallel for each component. For the main structure component with a high signal-to-noise ratio, a lightweight approach is used for feature extraction, reducing computational resource consumption. For the texture noise component with a low signal-to-noise ratio, noise feature filtering is performed to obtain the target texture features. This significantly improves the brightness of the original photoelectric signal image while avoiding noise increase to a certain extent, thus improving image quality.

Description

Technical Field

[0001] This disclosure relates to the field of image processing technology, specifically to an image processing method, apparatus, electronic device, and storage medium. Background Technology

[0002] During the generation of digital images, image signal processing (ISP) is required to convert the raw photoelectric signals captured by the sensor into a color image that can be recognized by the human eye. In low-light environments, the number of photons is small, resulting in a low signal-to-noise ratio of the raw photoelectric signals output by the sensor, leading to image distortion.

[0003] In related technologies, techniques such as Transformer enhancement are commonly used to perform self-attention calculations on the original photoelectric signal to optimize it. However, these optimization methods involve large computational loads and consume significant equipment resources. Summary of the Invention

[0004] This disclosure presents an image processing method, apparatus, electronic device, and storage medium.

[0005] The first aspect of this disclosure provides an image processing method, including: In response to acquiring the original photoelectric signal image, a frequency division operation is performed on the original photoelectric signal image to obtain the texture noise component and the main structure component; For the main structural component, extract the main structural features from the main structural component; and for the texture noise component, filter the noise in the texture noise component to obtain the target texture features. The target image is obtained by reconstructing the image based on the texture features and the main structural features.

[0006] In this embodiment of the disclosure, filtering noise from the texture noise component to obtain the target texture features includes: If i is less than n, the i-th texture feature output by the i-th downsampling module is used as the input of the (i+1)-th downsampling module to extract and downsample the i-th texture feature; if i is equal to n, the i-th texture feature is extracted to obtain the initial texture feature; where, if i=1, the input of the first downsampling module is the texture noise component; 1≤i≤n; n>1, n∈R; If j is less than n, then the j-th upsampled feature output by the j-th upsampled module and the (n-j+1)-th texture feature are used as the input of the (j+1)-th upsampled module for upsampling; if j equals n, then the j-th upsampled feature output by the j-th upsampled module is used for decoding to obtain the target texture feature; wherein, the input of the 1st upsampled module is the initial texture feature and the n-th texture feature output by the n-th downsampled module; 1≤j≤n.

[0007] In this embodiment of the disclosure, the feature extraction of the i-th texture feature includes: Global feature extraction is performed on the i-th texture feature to obtain the target global feature; The target global feature is concatenated with the i-th texture feature.

[0008] In this embodiment of the disclosure, the step of performing global feature extraction on the i-th texture feature to obtain the target global feature includes: Extract the local features of the i-th texture feature; Based on the contextual relationship of each element in the i-th texture feature, the initial global features of the texture noise component are extracted; The local features are multiplied by the initial global features to obtain the target global features.

[0009] In this embodiment of the disclosure, the step of performing a frequency division operation on the original photoelectric signal image to obtain a texture noise component and a main structure component includes: The original photoelectric signal image is converted into a frequency domain signal; Based on the frequency of the frequency domain signal, the frequency domain signal is divided into high-frequency components and low-frequency components; the high-frequency components are texture noise components; and the low-frequency components are main structure components.

[0010] In this embodiment of the disclosure, the original photoelectric signal image is frequency divided using any of the following methods: discrete wavelet transform, analytical filter method, or Laplace pyramid decomposition method.

[0011] In this embodiment of the disclosure, the extraction of main structural features from the main structural components includes: The main structural components are used to extract features using a convolutional neural network to obtain the main structural features.

[0012] A second aspect of this disclosure provides an image processing apparatus, the apparatus comprising: The frequency division module is used to perform a frequency division operation on the original photoelectric signal image in response to the acquisition of the original photoelectric signal image, to obtain the texture noise component and the main structure component. The feature extraction module is used to extract the main structural features from the main structural components; and to filter the noise in the texture noise components to obtain the target texture features. The reconstruction module is used to reconstruct the image based on the texture features and the main structural features to obtain the target image.

[0013] An embodiment of the third aspect of this disclosure provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described in the first aspect or any optional embodiment of the first aspect.

[0014] An embodiment of the fourth aspect of this disclosure provides a computer-readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the methods described in the first aspect and any optional embodiment of the first aspect.

[0015] The technical solutions provided in this disclosure have at least the following technical effects or advantages: In response to acquiring the original photoelectric signal image, a frequency division operation is performed on the original photoelectric signal image to obtain texture noise components and main structure components. By physically separating the texture noise and main structure, feature extraction is performed in parallel for the texture noise components and the main structure components respectively. For the main structure components with high signal-to-noise ratio, a lightweight method is used to extract the main structure features, reducing the consumption of computing resources. For the texture noise components with low signal-to-noise ratio, noise feature filtering is performed to obtain the target texture features. While significantly improving the brightness of the original photoelectric signal image, noise is avoided to a certain extent, thus improving the image quality.

[0016] Additional aspects and advantages of this disclosure will be set forth in part in the description which follows, and in part will be obvious from the description or may be learned by practice of this disclosure. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this disclosure. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart of an image processing method provided in an embodiment of this disclosure is shown; Figure 2 A schematic diagram of a denoising network in an image processing method provided in an embodiment of the present disclosure is shown; Figure 3This diagram illustrates the feature extraction of texture features in an image processing method provided by an embodiment of the present disclosure; Figure 4 This diagram illustrates global feature extraction in texture feature extraction according to an embodiment of the image processing method provided in this disclosure. Figure 5 A flowchart of an image processing apparatus provided according to an embodiment of the present disclosure is shown; Figure 6 A schematic diagram of the structure of an electronic device provided in an embodiment of the present disclosure is shown; Figure 7 A schematic diagram of a storage medium provided according to an embodiment of the present disclosure is shown. Detailed Implementation

[0018] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0019] It should be noted that, unless otherwise stated, the technical or scientific terms used in this disclosure shall have the ordinary meaning as understood by one of ordinary skill in the art to which this disclosure pertains.

[0020] During digital image generation, the raw photoelectric signals captured by the sensor are converted into a color image recognizable by the human eye through an image signal processing (ISP) pipeline. This process encompasses multiple stages, including black level correction, white balance, denoising, de-mosaicing, and color correction. In well-lit environments, a standard ISP pipeline can output high-quality images. However, in low-light or night vision environments, the scarcity of photons results in an extremely low signal-to-noise ratio (SNR) and a complex Poisson-Gaussian mixture of noise characteristics. Traditional ISP algorithms primarily rely on manually set fixed parameters or simple threshold rules for signal amplification and noise reduction. These fixed rules struggle to distinguish between extremely subtle texture details and random noise, often amplifying noise while simultaneously increasing brightness, leading to severe graininess, artifacts, and color distortion in the final image.

[0021] To overcome the limitations of traditional algorithms, deep learning-based image enhancement techniques have gradually become mainstream. Among them, Convolutional Neural Networks (CNNs) excel at handling local textures by extracting features through sliding windows on the image. However, since low-light enhancement tasks require not only restoring local details but also the overall brightness balance and color consistency of the image, CNNs, limited by their local receptive field, struggle to capture the relationships between distant pixels, easily leading to uneven brightness or glare in the enhanced image. To obtain global information, the Transformer architecture was introduced, thereby achieving a global receptive field.

[0022] The method of using the Transformer architecture to calculate the correlation between each pixel and all other pixels in an image through a self-attention mechanism can capture long-distance illumination dependencies to ensure global consistency. However, the computational cost of its illumination-guided self-attention module is proportional to the square of the image resolution. For 4K or even higher resolution images commonly found in mobile phones or surveillance cameras, the memory usage and computation time increase exponentially, making real-time processing difficult on edge devices with limited hardware resources.

[0023] This disclosure provides an image processing method, such as... Figure 1 The image shown is an embodiment of an image processing method provided by this disclosure, including the following steps: In step S11, in response to acquiring the original photoelectric signal image, a frequency division operation is performed on the original photoelectric signal image to obtain the texture noise component and the main structure component.

[0024] For example, a raw photoelectric signal image without complete image signal processing is acquired from an image sensor. This image is a two-dimensional matrix in the spatial domain, containing real-world scene signals and complex noise. A discrete wavelet basis (e.g., Haar wavelet) can be used to perform multi-level decomposition on the input raw image. This transformation converts the image from the spatial domain to the frequency domain. Unlike Fourier transforms, which tend to lose spatial location information, wavelet transforms can simultaneously preserve the temporal and spatial location and frequency domain features of the image. Through this transformation, the raw image is losslessly decomposed into four non-overlapping sub-bands: a low-frequency component containing global illumination distribution and main structure, and horizontal, vertical, and diagonal high-frequency components containing edge textures and random noise. This physical-level signal separation provides a solid data foundation for subsequently designing independent enhancement paths for different frequency characteristics.

[0025] The low-frequency approximate subband output by the discrete wavelet transform is defined as the "main structure component," which mainly contains information about the image's global illumination, large color regions, and the gradually changing contours of the main object. Meanwhile, the high-frequency detail subbands (typically corresponding to high-frequency information in the horizontal, vertical, and diagonal directions) are merged or processed separately and defined as the "texture noise component." This component is a collection of rapidly changing parts of the image, containing the object's true texture, edge details, and random noise such as shot noise and readout noise that needs to be removed.

[0026] In some embodiments, the above process can be summarized as follows: converting the original photoelectric signal image into a frequency domain signal; dividing the frequency domain signal into high-frequency components and low-frequency components based on the frequency of the frequency domain signal; the high-frequency components are texture noise components; and the low-frequency components are main structure components.

[0027] In addition to the discrete wavelet transform mentioned above, the texture noise component and the main structure component can also be obtained by using the analytical filter method or the Laplace pyramid decomposition method.

[0028] In step S12, for the main structure component, the main structure features are extracted from the main structure component; and for the texture noise component, the noise in the texture noise component is filtered to obtain the target texture features.

[0029] For example, regarding the main structural component, since it typically has a small spatial size at low resolution, computationally inexpensive structure-aware feature extraction algorithms can be employed. For instance, a Canny edge detector can be used to obtain precise boundaries; or low-frequency variants of local binary patterns can be computed to capture region shapes; or directional gradient histograms can be used to describe the directional distribution of the overall structure. Regarding the texture noise component, it is necessary to distinguish between real texture and random noise. Methods such as nonlocal mean denoising or wavelet thresholding can effectively remove Gaussian noise and other interference while preserving repetitive and structural texture patterns.

[0030] In other embodiments, the extraction of main structural features from the main structural components described above can be further enhanced by using convolutional neural networks (CNNs) to extract features from the main structural components, thereby obtaining the main structural features. Specifically, by utilizing the local receptive field features of the CNN, the grayscale histogram can be straightened and color balance restored. Since the resolution of the main structural components is relatively small compared to a portion of the original photoelectric signal image, the CNN can achieve a large receptive field coverage with extremely low computational cost, effectively capturing the global illumination distribution. Through end-to-end training, it can learn the mapping relationship from low-frequency signals in dark light to low-frequency signals in normal light, significantly improving image brightness while ensuring natural color transitions and avoiding the color banding phenomenon common in traditional methods.

[0031] In some embodiments, the process of filtering noise from the texture noise component to obtain the target texture features is problematic because the texture noise component contains both edge texture details and destructive random noise from the original photoelectric signal image. Traditional convolutional neural network methods, limited by their local receptive field, struggle to distinguish between edge texture details and random noise. Therefore, this application proposes a denoising network for the texture noise component to remove noise and obtain the target texture features. If i is less than n, the i-th texture feature output by the i-th downsampling module is used as the input of the (i+1)-th downsampling module to extract and downsample the i-th texture feature; if i equals n, the i-th texture feature is extracted to obtain the initial texture feature; where, if i=1, the input of the 1st downsampling module is the texture noise component; 1≤i≤n; n>1, n∈R; if j is less than n, the j-th upsampled feature output by the j-th upsampling module and the (n-j+1)-th texture feature are used as the input of the (j+1)-th upsampling module for upsampling; if j equals n, the j-th upsampled feature output by the j-th upsampling module is decoded to obtain the target texture feature; where, the input of the 1st upsampling module is the initial texture feature and the n-th texture feature output by the n-th downsampling module; 1≤j≤n.

[0032] For example, a denoising network such as Figure 2 As shown, the downsampling module and the feature extraction module are implemented in the same way when performing feature extraction. Compared with the feature extraction module, the downsampling module adds a downsampling process. Usually, after feature extraction and downsampling by i downsampling modules, the granularity of the feature is small enough, so only feature extraction is needed afterward. After obtaining the target texture feature through feature extraction, the target texture feature is upsampled to restore the target texture feature to the same size as the original photoelectric signal image.

[0033] To further prevent the loss of original high-frequency details during feature extraction in deep networks, this application introduces a skip connection mechanism between downsampling and upsampling modules at each level. Skip connections can directly pass features from previous layers to subsequent layers, forming a residual learning structure. This design not only solves the gradient vanishing problem in deep network training, but more importantly, residual connections allow the network to constantly reference the original high-frequency input information during denoising. This means the network only needs to learn and subtract the noise residual, without reconstructing the entire texture feature, thereby preserving the subtle edges and texture details of the image to the greatest extent and avoiding image blurring caused by over-denoising.

[0034] In some embodiments, feature extraction of the i-th texture feature includes: performing global feature extraction on the i-th texture feature to obtain a target global feature; and concatenating the target global feature with the i-th texture feature.

[0035] For example, global feature extraction refers to the process of mining and integrating the relationships and dependencies between all locations (pixels) of texture features, aiming to obtain a holistic and structured understanding of texture features.

[0036] After layer normalization, the input texture features are processed using the SS2D module (two-dimensional selective scanning module). The SS2D module scans and unfolds the normalized two-dimensional feature map into a one-dimensional sequence along multiple directions (such as top left to bottom right, top right to bottom left, etc.). Leveraging the linear complexity advantage of the state-space model, it establishes long-range dependencies between any two positions in the sequence, thereby capturing the contextual relationships across the entire feature map. Finally, the target global features are output through components such as a feedforward network. Skip connections are made between the target global features and texture features to achieve the fusion of multi-granularity information, ensuring that weak edge information is not smoothed.

[0037] like Figure 3 The diagram illustrates the feature extraction process described above. First, texture features are input into a global feature extraction module to obtain contextual relationships. This module includes the aforementioned SS2D module. The difference between the texture features and the target global features is calculated to obtain residual features, also known as local detail features. By subtracting the global background, subtle local variations in the original texture that were not explained by the global features are preserved, effectively removing low-frequency background interference while retaining high-frequency edges and details. Further, the residual features undergo feature fusion, convolution, and activation followed by attention masking to obtain the feature extraction result. Simultaneously, the response of noisy or flat regions is suppressed, making the feature extraction result more discriminative and selective.

[0038] The following describes the specific implementation process of global feature extraction for the target: extracting the local features of the i-th texture feature; extracting the initial global features of the texture noise component based on the contextual relationship of each element in the i-th texture feature; and multiplying the local features with the initial global features to obtain the target global features.

[0039] For example, the i-th texture feature is normalized by a layer and then input into a 3x3 depthwise separable convolutional layer. This convolution performs independent spatial convolution operations on each channel of the input feature map, limited to the 3x3 neighborhood, thereby capturing the local spatial patterns and microstructures at each location, such as basic elements like points, edges, and corners. Simultaneously, the i-th texture feature is input into the SS2D module. This module scans and unfolds the two-dimensional feature map into a one-dimensional sequence along multiple directions (such as the four diagonals). Utilizing a linear scanning mechanism, it establishes dependencies between any two distant locations during sequence propagation, thus modeling the long-range contextual relationships between all elements in the entire feature map. The local features are then multiplied element-wise with the initial global features. This is equivalent to reweighting the initial global features with local features (as spatially adaptive weights or gating signals). This achieves refined guidance of global semantics from local details. Regions with strong responses in the local features (potentially corresponding to important ground truth edges) enhance the weights of corresponding locations in the initial global features, thereby highlighting structured textures considered important at both the local and global levels.

[0040] like Figure 4 As shown, Figure 3 This diagram illustrates the global feature extraction process. After layer-level normalization of the texture features, the normalized texture features are processed through linear layers. One path is fed into the SS2D module, which efficiently captures long-range dependencies between features by scanning the feature map in multiple directions and applying a state-space model, extracting the initial global features. This output is then normalized to ensure numerical stability. The other path extracts local features (such as edges and texture details) through deep convolutional layers. The convolutional features are then subjected to an activation function to introduce non-linearity, enhancing the representational power of local features. The local features obtained from both paths are fused with the initial global features to achieve complementarity between the local and initial global features. Finally, a non-linear transformation is performed to obtain the target global features.

[0041] In step S13, the image is reconstructed based on texture features and main structure features to obtain the target image.

[0042] For example, the reconstruction process transforms texture features and main structural features into a high-resolution image space recognizable by the human eye, forming a closed-loop link in the entire image processing process. The complete reconstruction characteristic of wavelet transform can be utilized by employing a corresponding discrete wavelet inverse transform algorithm. In practice, the main structural features and texture features are first received, and the texture features are re-decomposed into corresponding horizontal, vertical, and diagonal sub-bands, with strict dimensional alignment. Subsequently, these frequency bands are fused using the inverse transform operator, mathematically ensuring lossless signal restoration except for enhancement processing. This process not only reconstructs the spatial resolution of the image but also organically integrates the enhanced illumination information with the denoised detail information, ultimately outputting a high-quality low-light image with uniform brightness, extremely low noise, and clear texture, effectively avoiding the block effect or halo phenomenon that may occur during direct synthesis in the spatial domain.

[0043] The image processing method of this application embodiment, in response to acquiring the original photoelectric signal image, performs a frequency division operation on the original photoelectric signal image to obtain a texture noise component and a main structure component. By physically separating the texture noise and the main structure, feature extraction is performed in parallel for the texture noise component and the main structure component respectively. For the main structure component with a high signal-to-noise ratio, a lightweight method is used to extract the main structure features, reducing the computational resource consumption. For the texture noise component with a low signal-to-noise ratio, noise feature filtering is performed to obtain the target texture features. While significantly improving the brightness of the original photoelectric signal image, it avoids increasing noise to a certain extent and improves the image quality.

[0044] The image processing method of this application embodiment balances computational efficiency and image imaging quality. It efficiently recovers illumination using the main structure component and accurately removes noise while preserving texture using the texture noise component. This not only achieves adaptive enhancement of low-light images but also effectively reduces computational complexity, enabling it to be deployed on resource-constrained hardware devices, and significantly improves the overall quality and practicality of night vision imaging.

[0045] Existing techniques for directly brightening images in the spatial domain often result in the simultaneous amplification of noise in dark areas. This application uses wavelet transform to physically separate illumination information (subject structure features) from noise information (texture noise components), performing significant brightening only in the low-frequency domain while focusing on noise reduction in the high-frequency domain. This effectively resolves the contradiction between noise amplification and detail loss while significantly improving image brightness, avoiding the emergence of grainy noise and ensuring image purity. This approach combines skip connections and global feature extraction in the process of dealing with noise information (texture noise components). The selective scanning mechanism can keenly capture the structured features of the texture, while the skip connection mechanism ensures that the original high-frequency details are not lost as the network depth increases. The combination of these two mechanisms solves the problem that traditional convolutional networks easily misjudge weak textures as noise and "smooth them out," allowing the enhanced image to retain rich edge details while denoising.

[0046] Corresponding to the above implementation methods, this disclosure also provides an image processing apparatus for performing the above-described methods. Figure 1 The image processing method of any of the illustrated embodiments, such as Figure 5 As shown, the image processing apparatus includes: Frequency division module 501 is used to perform frequency division operation on the original photoelectric signal image in response to acquiring the original photoelectric signal image, to obtain texture noise component and main structure component; The feature extraction module 502 is used to extract the main structure features from the main structure component; and to filter the noise in the texture noise component to obtain the target texture features. The reconstruction module 503 is used to reconstruct the image based on the texture features and the main structural features to obtain the target image.

[0047] Optionally, the feature extraction module is further configured to: If i is less than n, the i-th texture feature output by the i-th downsampling module is used as the input of the (i+1)-th downsampling module to extract and downsample the i-th texture feature; if i is equal to n, the i-th texture feature is extracted to obtain the initial texture feature; where, if i=1, the input of the first downsampling module is the texture noise component; 1≤i≤n; n>1, n∈R; If j is less than n, then the j-th upsampled feature output by the j-th upsampled module and the (n-j+1)-th texture feature are used as the input of the (j+1)-th upsampled module for upsampling; if j equals n, then the j-th upsampled feature output by the j-th upsampled module is used for decoding to obtain the target texture feature; wherein, the input of the 1st upsampled module is the initial texture feature and the n-th texture feature output by the n-th downsampled module; 1≤j≤n.

[0048] Optionally, the feature extraction module is further configured to: perform global feature extraction on the i-th texture feature to obtain the target global feature; The target global feature is concatenated with the i-th texture feature.

[0049] Optionally, the feature extraction module is further configured to: extract local features of the i-th texture feature; Based on the contextual relationship of each element in the i-th texture feature, the initial global features of the texture noise component are extracted; The local features are multiplied by the initial global features to obtain the target global features.

[0050] Optionally, the frequency division module is further configured to: convert the original photoelectric signal image into a frequency domain signal; Based on the frequency of the frequency domain signal, the frequency domain signal is divided into high-frequency components and low-frequency components; the high-frequency components are texture noise components; and the low-frequency components are main structure components.

[0051] Optionally, the original photoelectric signal image may be frequency divided using any of the following methods: discrete wavelet transform, analytical filter, or Laplace pyramid decomposition.

[0052] Optionally, a convolutional neural network is used to extract features from the main structural components to obtain the main structural features.

[0053] The image processing apparatus and image processing method provided in the above embodiments of this disclosure are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the applications stored therein.

[0054] This disclosure also provides an electronic device for performing the above-described method. Please refer to... Figure 6 This illustrates a schematic diagram of an electronic device provided by some embodiments of the present disclosure. For example... Figure 6 As shown, the electronic device includes: a processor 600, a memory 601, a bus 602, and a communication interface 603. The processor 600, the communication interface 603, and the memory 601 are connected via the bus 602. The memory 601 stores a computer program that can run on the processor 600. When the processor 600 runs the computer program, it executes the aforementioned provisions of this disclosure. Figure 1 The method provided by any of the illustrated embodiments.

[0055] The memory 601 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 603 (which can be wired or wireless), such as the Internet, wide area network, local area network, or metropolitan area network.

[0056] Bus 602 can be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. Memory 601 is used to store programs, and the processor 600 executes the programs after receiving execution instructions. Figure 1 The illustrated method can be applied to or implemented by the processor 600.

[0057] The processor 600 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of the processor 600 or by instructions in software form. The processor 600 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this disclosure. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this disclosure can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules may reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory 601. Processor 600 reads the information in memory 601 and, in conjunction with its hardware, completes the steps of the above method.

[0058] The electronic devices and methods provided in this disclosure are based on the same inventive concept and have the same beneficial effects as the methods they employ, operate, or implement.

[0059] This disclosure also provides a computer-readable storage medium corresponding to the method provided in the foregoing embodiments. Please refer to... Figure 7 The computer-readable storage medium shown is an optical disc 30, on which a computer program (i.e., a program product) is stored, which, when run by a processor, executes the methods provided in any of the foregoing embodiments.

[0060] It should be noted that examples of the computer-readable storage medium may also include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other optical and magnetic storage media, which will not be elaborated here.

[0061] The computer-readable storage medium provided in the above embodiments of this disclosure and the methods provided in the embodiments of this disclosure are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the applications stored therein.

[0062] It should be noted that: Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of this disclosure may be practiced without these specific details. In some instances, well-known structures and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0063] Similarly, it should be understood that, in order to simplify this disclosure and aid in understanding one or more of the various inventive aspects, in the foregoing description of exemplary embodiments of this disclosure, various features of this disclosure are sometimes grouped together in a single embodiment, figure, or description thereof. However, this approach to disclosure should not be construed as reflecting a schematic diagram in which the claimed disclosure requires more features than are explicitly stated in each embodiment. The inventive aspect lies in fewer than all features of a single foregoing disclosed embodiment. Therefore, the implementation following the detailed description is hereby explicitly incorporated into this detailed description, wherein each embodiment itself serves as a separate embodiment of this disclosure.

[0064] Furthermore, those skilled in the art will understand that although some embodiments described herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this disclosure and form different embodiments.

[0065] The above description is merely a preferred embodiment of this disclosure, but the scope of protection of this disclosure is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this disclosure should be included within the scope of protection of this disclosure.

Claims

1. An image processing method, characterized in that, The method includes: In response to acquiring the original photoelectric signal image, a frequency division operation is performed on the original photoelectric signal image to obtain the texture noise component and the main structure component; For the main structural component, extract the main structural features from the main structural component; and for the texture noise component, filter the noise in the texture noise component to obtain the target texture features. The target image is obtained by reconstructing the image based on the texture features and the main structural features.

2. The method according to claim 1, characterized in that, The step of filtering noise from the texture noise component to obtain the target texture features includes: If i is less than n, the i-th texture feature output by the i-th downsampling module is used as the input of the (i+1)-th downsampling module to extract and downsample the i-th texture feature; if i is equal to n, the i-th texture feature is extracted to obtain the initial texture feature; where, if i=1, the input of the first downsampling module is the texture noise component; 1≤i≤n; n>1, n∈R; If j is less than n, then the j-th upsampled feature output by the j-th upsampled module and the (n-j+1)-th texture feature are used as the input of the (j+1)-th upsampled module for upsampling; if j equals n, then the j-th upsampled feature output by the j-th upsampled module is used for decoding to obtain the target texture feature; wherein, the input of the 1st upsampled module is the initial texture feature and the n-th texture feature output by the n-th downsampled module; 1≤j≤n.

3. The method according to claim 2, characterized in that, The feature extraction of the i-th texture feature includes: Global feature extraction is performed on the i-th texture feature to obtain the target global feature; The target global feature is concatenated with the i-th texture feature.

4. The method according to claim 3, characterized in that, The step of performing global feature extraction on the i-th texture feature to obtain the target global feature includes: Extract the local features of the i-th texture feature; Based on the contextual relationship of each element in the i-th texture feature, the initial global features of the texture noise component are extracted; The local features are multiplied by the initial global features to obtain the target global features.

5. The method according to any one of claims 1-4, characterized in that, The step of performing a frequency division operation on the original photoelectric signal image to obtain texture noise components and main structure components includes: The original photoelectric signal image is converted into a frequency domain signal; Based on the frequency of the frequency domain signal, the frequency domain signal is divided into high-frequency components and low-frequency components; the high-frequency components are texture noise components; and the low-frequency components are main structure components.

6. The method according to claim 5, characterized in that, The original photoelectric signal image is subjected to frequency division using any of the following methods: discrete wavelet transform, analytical filter method, or Laplace pyramid decomposition method.

7. The method according to any one of claims 1-4, characterized in that, The extraction of main structural features from the main structural components includes: The main structural components are used to extract features using a convolutional neural network to obtain the main structural features.

8. An image processing apparatus, characterized in that, The device includes: The frequency division module is used to perform a frequency division operation on the original photoelectric signal image in response to the acquisition of the original photoelectric signal image, to obtain the texture noise component and the main structure component. The feature extraction module is used to extract the main structural features from the main structural components; and to filter the noise in the texture noise components to obtain the target texture features. The reconstruction module is used to reconstruct the image based on the texture features and the main structural features to obtain the target image.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory, wherein the processor, when executing the computer program, implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program that, when executed by a processor, implements the method of any one of claims 1 to 7.

Images