Image noise level determination method and device, readable storage medium and program product

Abstract

The invention relates to an image noise level determination method and device, computer equipment, a computer readable storage medium and a computer program product. The method comprises the following steps: performing feature extraction on a target image to obtain image features of the target image; performing noise prediction based on the image features to obtain a noise map corresponding to the target image; the noise map describes noise intensity corresponding to each pixel point in the target image; based on the semaphore corresponding to each pixel point in the target image, adjusting the noise intensity of the corresponding pixel point in the noise map to obtain an adjusted noise map; and determining a noise level corresponding to each pixel point based on the adjusted noise intensity corresponding to each pixel point in the target image described by the adjusted noise map. By adopting the method, the accuracy of image noise evaluation can be improved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method, apparatus, readable storage medium, and program product for determining image noise levels. Background Technology

[0002] With the development of the internet and multimedia technologies, massive amounts of images are widely collected, stored, and disseminated in scenarios such as social media, security monitoring, medical imaging, and industrial inspection. Image quality directly affects the effectiveness of downstream tasks. Therefore, how to accurately evaluate and analyze image quality, especially noise levels, has become a crucial research issue in the field of image processing.

[0003] Currently, image noise assessment typically relies on an overall quality score or a specific noise model, which often only outputs a single value, thus failing to accurately assess the noise level of an image. Summary of the Invention

[0004] Based on this, this application provides a method, apparatus, computer device, computer-readable storage medium, and computer program product for determining image noise levels, which can improve the accuracy of image noise assessment.

[0005] On the one hand, this application provides a method for determining image noise level, including:

[0006] Feature extraction is performed on the target image to obtain its image features;

[0007] Based on the image features, noise prediction is performed to obtain a noise map corresponding to the target image; the noise map describes the noise intensity corresponding to each pixel in the target image.

[0008] Based on the signal quantity corresponding to each pixel in the target image, the noise intensity of the corresponding pixel in the noise map is adjusted to obtain the adjusted noise map;

[0009] Based on the adjusted noise intensity corresponding to each pixel in the target image described by the adjusted noise map, the noise level corresponding to each pixel is determined.

[0010] In one embodiment, the image features include multi-scale features and semantic features; the feature extraction of the target image to obtain the image features of the target image includes:

[0011] The target image is feature extracted by the encoder of the noise prediction model to obtain multi-scale features and semantic features;

[0012] The step of performing noise prediction based on the image features to obtain a noise map corresponding to the target image includes:

[0013] The multi-scale features and semantic features are fused using the decoder of the noise prediction model to obtain fused features;

[0014] Based on the fusion features, noise prediction is performed to obtain the noise map corresponding to the target image.

[0015] In one embodiment, the method further includes:

[0016] Obtain the sample image and the corresponding label noise map of the sample image;

[0017] The encoder of the initial noise prediction model extracts features from the sample images to obtain multi-scale sample features and sample semantic features;

[0018] The multi-scale sample features and the sample semantic features are fused by the decoder of the initial noise prediction model to obtain the sample fusion features;

[0019] Based on the sample fusion features, noise prediction is performed to obtain a sample noise map of the sample image;

[0020] Based on the sample noise map and the label noise map, the parameters of the initial noise prediction model are optimized to obtain a trained noise prediction model.

[0021] In one embodiment, obtaining the sample image and the corresponding label noise map includes:

[0022] Obtain high-resolution raw images;

[0023] The original image is subjected to a first degradation process to obtain a sample image;

[0024] The original image is subjected to a second degradation process to obtain a degraded reference image; the first degradation process includes the second degradation process and noise addition processing.

[0025] A label noise map is determined based on the degraded reference map and the sample image.

[0026] In one embodiment, determining the label noise map based on the degraded reference map and the sample image includes:

[0027] Based on the degree of similarity between the degraded reference image and the sample image, a first similarity map and a second similarity map are determined;

[0028] The first similarity map and the second similarity map are merged to obtain a merged similarity map;

[0029] The label noise map is determined based on the fused similarity distribution map.

[0030] In one embodiment, the step of optimizing the parameters of the initial noise prediction model based on the sample noise map and the label noise map to obtain a trained noise prediction model includes:

[0031] The training loss value is determined based on the sample noise map and the label noise map;

[0032] Based on the training loss value, the encoder and decoder of the initial noise prediction model are optimized until the convergence condition is met, thus obtaining a trained noise prediction model.

[0033] In one embodiment, the method further includes:

[0034] The target image is semantically segmented to obtain a segmented semantic map;

[0035] Based on the segmentation semantic map and the adjusted noise map, the regions containing noise in the target image are located to obtain a noise localization map.

[0036] On the one hand, this application also provides an image noise level determination device, including:

[0037] The image feature extraction module is used to extract features from the target image to obtain the image features of the target image;

[0038] The noise map acquisition module is used to predict noise based on the image features to obtain a noise map corresponding to the target image; the noise map describes the noise intensity corresponding to each pixel in the target image.

[0039] The noise map adjustment module is used to adjust the noise intensity of corresponding pixels in the noise map based on the signal quantity corresponding to each pixel in the target image, so as to obtain the adjusted noise map.

[0040] The noise level determination module is used to determine the noise level corresponding to each pixel based on the adjusted noise intensity corresponding to each pixel in the target image described by the adjusted noise map.

[0041] On the one hand, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods described above.

[0042] In one aspect, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0043] On the one hand, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of any of the methods described above.

[0044] The aforementioned image noise level determination method, apparatus, computer equipment, computer-readable storage medium, and computer program product extract features from the target image to obtain image features. Based on these image features, noise prediction is performed to obtain a noise map corresponding to the target image. The noise map describes the noise intensity corresponding to each pixel in the target image, comprehensively reflecting the noise distribution of each pixel in the image. Based on the signal quantity corresponding to each pixel in the target image, the noise intensity of the corresponding pixel in the noise map is adjusted to obtain an adjusted noise map. Introducing signal quantity to adjust the noise intensity allows the noise prediction result to be corrected by combining local brightness and texture characteristics, making it more consistent with the subjective perception of noise by the human eye. The resulting adjusted noise map not only improves the accuracy and robustness of noise estimation but also effectively distinguishes the salience of noise in different regions. Based on the adjusted noise intensity corresponding to each pixel in the target image described by the adjusted noise map, the noise level corresponding to each pixel is determined, realizing fine-grained quantitative classification of image noise, thereby improving the accuracy of image noise assessment. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is an application environment diagram of the image noise level determination method in one embodiment;

[0047] Figure 2 This is a flowchart illustrating an image noise level determination method in one embodiment;

[0048] Figure 3 This is a schematic diagram of the structure of a noise prediction model in one embodiment;

[0049] Figure 4 This is a flowchart illustrating the image noise level determination method in another embodiment;

[0050] Figure 5 This is a flowchart illustrating the image noise level determination method in another embodiment;

[0051] Figure 6 This is a structural block diagram of an image noise level determination device in one embodiment;

[0052] Figure 7 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0053] To make the objectives, technical solutions, and beneficial effects of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0054] The image noise level determination method provided in this application embodiment can be applied to, for example, Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated on server 104 or placed in the cloud or on another network server. This image noise level determination method can be executed by terminal 102 or server 104 alone, or by terminal 102 and server 104 collaboratively. In some embodiments, the image noise level determination method is executed by terminal 102. Terminal 102 extracts features from the target image to obtain image features; performs noise prediction based on the image features to obtain a noise map corresponding to the target image; the noise map describes the noise intensity corresponding to each pixel in the target image; adjusts the noise intensity of the corresponding pixel in the noise map based on the signal quantity corresponding to each pixel in the target image to obtain an adjusted noise map; and determines the noise level corresponding to each pixel based on the adjusted noise intensity described by the adjusted noise map.

[0055] The terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle systems, projection devices, skin detection devices, etc. Portable wearable devices can include smartwatches, smart bracelets, head-mounted displays, etc. Head-mounted displays can be virtual reality (VR) devices, augmented reality (AR) devices, smart glasses, etc. The server 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0056] In one embodiment, such as Figure 2 As shown, an image noise level determination method is provided, which can be applied to... Figure 1Taking a computer device (terminal 102 or server 104) as an example, the method may include the following steps:

[0057] S202, extract features from the target image to obtain the image features of the target image.

[0058] The target image refers to the image to be tested that requires noise level determination, and it can originate from various scenarios such as social media uploads, surveillance data collection, medical image scanning, or industrial inspection. For example, the terminal can display an interactive page with an image upload portal, through which users can upload target images for which noise level determination is required.

[0059] Feature extraction refers to processing a target image using a pre-defined convolutional neural network encoder to extract its representation information at different levels and scales. This process includes not only low-level local features such as edges and textures, but also high-level semantic information and global context, thus providing multi-scale and multi-semantic feature support for subsequent noise prediction.

[0060] Image features refer to the set of multi-scale features and semantic features obtained from the feature extraction process. Multi-scale features are used to describe the changes in detail at different resolutions, while semantic features are used to characterize the overall structure and contextual patterns. These image features, as intermediate representations, can provide the decoder with the necessary information to output pixel-level noise prediction results.

[0061] Specifically, the computer equipment acquires the target image whose noise level needs to be determined, and inputs the target image into a pre-trained noise prediction model. The noise prediction model then processes the target image to obtain its image features.

[0062] Among them, the noise prediction model refers to a deep neural network model that is pre-trained on a large-scale image dataset. This large-scale image dataset includes high-resolution original images and sample images obtained through degradation processing. The high-resolution original images can specifically come from large-scale high-resolution image datasets (such as DF2K, LSDIR, etc.).

[0063] It should be noted that the noise prediction model provided in this application embodiment can specifically be a deep convolutional neural network model with an encoder-decoder structure. The encoder is used to perform multi-layer convolution operations and feature extraction on the input image to obtain image features.

[0064] In one embodiment, the computer device may also perform size normalization processing on the target image to be processed, adjust the target image to an adjusted image of the target size, and then input the adjusted image of the target size into a pre-trained noise prediction model.

[0065] The target size refers to a uniform resolution, such as 512×512 pixels, that is pre-set according to the input requirements of the noise prediction model. This is used to ensure that input images of different resolutions can maintain a consistent feature scale and computational complexity when entering the model. The adjusted image refers to the image data that matches the target size after interpolating and scaling the original target image. The adjusted image maintains a correspondence with the original target image in terms of spatial structure, thereby ensuring the spatial consistency of noise distribution during model inference.

[0066] S204. Based on image features, noise prediction is performed to obtain a noise map corresponding to the target image; the noise map describes the noise intensity corresponding to each pixel in the target image.

[0067] The noise map is a two-dimensional image matrix generated by the noise prediction process that maintains a one-to-one correspondence with the target image in terms of spatial resolution. Each pixel value in the noise map is used to characterize the noise intensity of the corresponding pixel in the target image, and its value reflects the degree to which the pixel is affected by noise.

[0068] Specifically, after obtaining the image features of the target image, the computer equipment processes the image features of the target image through a noise prediction model to obtain the noise map corresponding to the target image.

[0069] Specifically, the noise prediction model can be a deep convolutional neural network model with an encoder-decoder structure. The decoder is used to process the image features extracted by the encoder to output a noise map describing the noise intensity of each pixel in the target image.

[0070] S206, Based on the signal quantity corresponding to each pixel in the target image, adjust the noise intensity of the corresponding pixel in the noise map to obtain the adjusted noise map.

[0071] The semaphore, calculated by combining the brightness and texture features of the target image, is an indicator used to characterize the visual masking effect of pixels. Specifically, brightness features reflect the human eye's sensitivity to noise in different bright and dark areas, while texture features reflect the differences in the human eye's perception of noise in complex or smooth areas. The semaphore value is obtained by fusing the mean brightness, variance brightness, local gradient, and local entropy of each pixel within its neighborhood. A larger semaphore indicates a stronger visual masking ability in the region where the pixel is located, and a weaker impact of noise on the human eye; a smaller semaphore indicates a weaker visual masking ability, and a more significant impact of noise on the human eye.

[0072] Specifically, after obtaining the noise map, the computer device can further acquire the signal quantity corresponding to each pixel in the target image, and adjust the noise intensity of the corresponding pixel in the noise map based on the signal quantity corresponding to each pixel to obtain the adjusted noise intensity of each pixel, and generate an adjusted noise map based on the adjusted noise intensity of each pixel.

[0073] In one embodiment, the computer device acquires the signal quantity corresponding to each pixel in the target image, and for any pixel, determines the signal-to-noise ratio value corresponding to the pixel based on the signal quantity and noise intensity, and determines the adjusted noise intensity corresponding to the pixel based on the signal-to-noise ratio value.

[0074] In one embodiment, the computer device may determine the adjusted noise intensity for any given pixel in the following manner:

[0075]

[0076]

[0077] Where 𝑥 represents any pixel in the target image; This represents the signal quantity corresponding to the pixel point calculated based on the brightness and texture features of the target image; This indicates the noise intensity of that pixel in the noise map; This represents the signal-to-noise ratio of the pixel, used to measure the perceptible noise level of the pixel under local brightness and texture conditions. This indicates the adjusted noise intensity, the magnitude of which reflects the visual significance of the noise at that pixel.

[0078] Because the adjusted noise map is based on the noise map by incorporating the signal quantities corresponding to each pixel for weighted adjustment, the noise intensity of each pixel not only reflects its objective noise magnitude but also incorporates the visual perceptibility under local brightness and texture conditions. Therefore, the adjusted noise map can comprehensively reflect the relative relationship between noise and signal, thus better conforming to the subjective perception of the human eye. Based on this characteristic, the adjusted noise map can also be called a signal-to-noise ratio map, and its core function is to transform the physical noise intensity into a noise saliency distribution that conforms to the laws of visual perception.

[0079] S208, Based on the adjusted noise intensity corresponding to each pixel in the target image described by the adjusted noise map, determine the noise level corresponding to each pixel.

[0080] The noise level refers to the discrete levels divided according to the adjusted noise intensity of each pixel in the adjusted noise map, used to characterize the visual significance of noise at that pixel. Specifically, the noise level can be determined based on a preset threshold range or statistical methods (such as quantile division, mean-variance analysis), mapping continuous noise intensity values ​​to several levels, such as five levels, corresponding to extremely low, relatively low, medium, relatively high, and extremely high noise levels, respectively.

[0081] Specifically, after obtaining the adjusted noise map, the computer device can obtain the corresponding adjusted noise intensity for each pixel in the adjusted noise map, and compare the adjusted noise intensity with the preset noise level classification rules to determine the noise level corresponding to each pixel.

[0082] In one embodiment, the computer device can further divide the target image into regions to obtain different image sub-regions, and determine the noise level corresponding to each image sub-region based on the adjusted noise intensity corresponding to the pixels in each image sub-region.

[0083] For example, when the target image is a portrait photograph, the image can be divided into "face regions" and "background regions." The adjusted noise intensity of pixels in the face region is statistically analyzed; if the average or quantile falls within a higher range, the noise level of the face region is determined to be high, while the background region may correspond to a lower noise level. Similarly, when the target image is a scene image, it can be divided into "sky regions," "building regions," and "vegetation regions" based on semantic segmentation results, and the noise level of each region can be determined separately, thus achieving regionalized noise analysis that better meets the needs of practical applications.

[0084] In the above-described method for determining image noise levels, feature extraction is performed on the target image to obtain its image features. Noise prediction is then performed based on these features to obtain a noise map corresponding to the target image. This noise map describes the noise intensity of each pixel in the target image, comprehensively reflecting the noise distribution of each pixel. Based on the semaphore corresponding to each pixel in the target image, the noise intensity of the corresponding pixel in the noise map is adjusted to obtain an adjusted noise map. Introducing semaphore to adjust the noise intensity allows the noise prediction results to be corrected by incorporating local brightness and texture characteristics, making it more consistent with the subjective perception of noise by the human eye. The resulting adjusted noise map not only improves the accuracy and robustness of noise estimation but also effectively distinguishes the salience of noise in different regions. Based on the adjusted noise intensity described by the adjusted noise map for each pixel in the target image, the noise level corresponding to each pixel is determined, achieving fine-grained quantitative classification of image noise and thus improving the accuracy of image noise assessment.

[0085] In one embodiment, the image features include multi-scale features and semantic features; the process by which the computer device extracts features from the target image to obtain the image features of the target image includes: extracting features from the target image through the encoder of the noise prediction model to obtain multi-scale features and semantic features; the process by which the computer device performs noise prediction based on the image features to obtain the noise map corresponding to the target image includes: fusing the multi-scale features and semantic features through the decoder of the noise prediction model to obtain fused features; and performing noise prediction based on the fused features to obtain the noise map corresponding to the target image.

[0086] The encoder refers to a neural network structure used to perform layer-by-layer convolution operations and downsampling on the input target image. It extracts feature representations at different levels through multiple convolution kernels. In this embodiment, the encoder can be constructed based on a convolutional neural network (ConvNet) structure.

[0087] Multi-scale features refer to feature maps extracted by the encoder during different convolutions and downsampling processes in the low-level network, which are used to characterize the feature information of the target image at the level of detail and in local regions.

[0088] Semantic features refer to feature representations extracted by the encoder in a high-level network that contain global context and abstract semantics, and are used to reflect information about the target image at the level of overall structure and category.

[0089] A decoder is a neural network structure used to fuse multi-scale features and semantic features extracted by the encoder, and to restore the feature map to the same spatial resolution as the target image through progressive upsampling and convolution operations, so as to output a noise map representing the noise intensity of each pixel in the target image. In an optional embodiment, the decoder can be built based on an improved DeeplabV3 structure, which expands the receptive field through dilated convolution to enhance the global feature acquisition capability, and combines skip connections and residual connections to achieve effective fusion of features at different scales.

[0090] Fusion features refer to the feature representation obtained by fusing multi-scale features and semantic features from the encoder in the decoder. Fusion features have both local detail information and global semantic information, which can more accurately represent the noise distribution in different regions of the target image and provide high-quality feature input for generating noise maps.

[0091] Specifically, after acquiring the target image, the computer device inputs it into the encoder of a pre-trained noise prediction model. The encoder sequentially performs layer-by-layer convolution operations and downsampling on the target image, extracting low-level features such as edges, textures, and local details at multiple scales using different convolution kernels, thus obtaining multi-scale features to represent the local information of the target image. Simultaneously, in the high-level network, the encoder extracts high-level features containing global contextual information and abstract semantics through deeper convolutions and nonlinear mappings, obtaining semantic features to represent the overall structure and category information of the target image. After obtaining the multi-scale and semantic features, the computer device then processes the multi-scale... The feature and semantic features are input into the decoder of the noise prediction model. The decoder fuses features at different levels through skip connections and residual connections to obtain fused features that combine local details and global semantics. Then, the fused features are sequentially upsampled and convolutionally processed to gradually restore their spatial resolution to be consistent with the target image. In this process, the decoder expands the receptive field through dilated convolution to capture the global contextual information of the image and enhances the feature representation ability by combining nonlinear activation functions. Finally, the decoder performs pixel-level regression on the restored high-resolution feature map using the prediction layer and outputs a noise map that corresponds one-to-one with the target image in terms of spatial resolution.

[0092] like Figure 3 The diagram shows a structural schematic of a noise prediction model in one embodiment. The noise prediction model adopts an encoder-decoder structure, which consists of an encoder based on a convolutional neural network (ConvNet) and a decoder based on an improved version of DeeplabV3. The encoder consists of multiple convolutional blocks and downsampling units. The input target image first undergoes layer-by-layer convolution and downsampling operations from convolutional block 1 to convolutional block 4 to extract low-level features such as edges, textures, and local details at different scales. Simultaneously, semantic features containing global contextual information and abstract semantics are obtained in deeper layers of the network. This preserves local detail features while gaining a global understanding of the overall structure. After receiving the initial fused features obtained from the multi-scale feature fusion, the decoder first performs dimensionality reduction and feature compression on the initial fused features through 1×1 convolution. The processing result is then concatenated with the semantic features output from higher layers of the encoder to obtain enhanced fused features. Subsequently, the decoder gradually restores the spatial resolution of the feature map using progressive upsampling and 3×3 convolution operations. Shallow features are introduced into the upsampling process through skip connections to supplement detail information and edge structures. During this process, the decoder expands the receptive field through dilated convolution to obtain global contextual information of the image and combines nonlinear activation functions to enhance feature representation capabilities. Finally, the decoder performs pixel-level regression on the restored high-resolution feature map using prediction layers, outputting a noise map that corresponds one-to-one with the target image in terms of spatial resolution.

[0093] In the above embodiments, the computer device extracts features from the target image through the encoder of the noise prediction model to obtain multi-scale features and semantic features. The multi-scale features and semantic features are then fused through the decoder of the noise prediction model to obtain fused features. Based on the fused features, noise prediction is performed to obtain the noise map corresponding to the target image. This achieves pixel-level fine prediction of noise in the target image, so that the noise intensity of each pixel in the target image can be accurately characterized, thereby significantly improving the accuracy and interpretability of noise assessment.

[0094] In one embodiment, the above-mentioned image noise level determination method further includes a process of training a noise prediction model, which specifically includes the following steps: acquiring a sample image and a corresponding label noise map; extracting features from the sample image using the encoder of the initial noise prediction model to obtain multi-scale sample features and sample semantic features; fusing the multi-scale sample features and sample semantic features using the decoder of the initial noise prediction model to obtain sample fusion features; performing noise prediction based on the sample fusion features to obtain a sample noise map of the sample image; and optimizing the parameters of the initial noise prediction model based on the sample noise map and the label noise map to obtain a trained noise prediction model.

[0095] Among them, sample images refer to training data images containing noise used for model training. Sample images can be generated from high-definition original images through degradation processing. Degradation processing includes, but is not limited to, one or more operations such as blurring, resolution compression, noise addition, compression encoding or video compression, to simulate various noise degradation situations that may occur in real application scenarios. High-definition original images can specifically come from large-scale high-definition image datasets (such as DF2K, LSDIR, etc.).

[0096] The label noise map corresponding to the sample image is a supervision signal image that corresponds one-to-one with the sample image in terms of spatial resolution, used to characterize the true noise intensity of each pixel in the sample image.

[0097] Specifically, the computer acquires sample images and corresponding label noise maps at spatial resolution. The sample images are input into the encoder of the initial noise prediction model. The encoder extracts multi-scale sample features and sample semantic features through layer-by-layer convolution and downsampling operations. These features are then input into the decoder of the initial noise prediction model. The decoder fuses the multi-scale and semantic features to obtain fused sample features that combine local details and global semantics. Based on these fused features, it performs upsampling and convolution operations at each level. After restoring the model to the same spatial resolution as the sample images, the prediction layer outputs pixel-level regression results to obtain sample noise maps that correspond one-to-one with the sample images. The sample noise maps are compared with the label noise maps, and the parameters of the initial noise prediction model are optimized and updated based on the differences between them. This training process is repeated until the convergence condition is met, resulting in a trained noise prediction model.

[0098] In the above embodiments, the computer device acquires a sample image and its corresponding label noise map, extracts features from the sample image using the encoder of the initial noise prediction model to obtain multi-scale sample features and sample semantic features, fuses the multi-scale sample features and sample semantic features using the decoder of the initial noise prediction model to obtain sample fusion features, performs noise prediction based on the sample fusion features to obtain a sample noise map of the sample image, and optimizes the parameters of the initial noise prediction model based on the sample noise map and the label noise map to obtain a trained noise prediction model. This enables the noise prediction model to gradually learn noise features under different scenarios and intensities during training, improving the accuracy and robustness of the model in pixel-level noise prediction, and providing a reliable guarantee for subsequent refined noise intensity estimation and level classification on the target image.

[0099] In one embodiment, the process of a computer device acquiring a sample image and a corresponding label noise map includes the following steps: acquiring a high-resolution original image, performing a first degradation process on the original image to obtain a sample image; performing a second degradation process on the original image to obtain a degradation reference image; the first degradation process includes the second degradation process and noise addition process, and determining the label noise map based on the degradation reference image and the sample image.

[0100] The second degradation process refers to an image degradation operation performed on the high-definition original image without adding noise. The degradation operation may include one or more of the following: blurring, resolution compression, compression encoding, or video compression.

[0101] The first degradation process refers to the image degradation process that further performs noise addition operations on the basis of the second degradation process on the high-definition original image; the noise addition may include one or more of Gaussian noise, salt and pepper noise, color noise or compressed noise.

[0102] Specifically, the computer equipment first acquires a high-definition original image, performs a first degradation process on the original image to obtain a sample image containing noise; then, it performs a second degradation process on the original image alone to obtain a degradation reference image that has the same degradation features as the sample image but does not contain noise. After that, the degradation reference image and the sample image are measured at the pixel level, and the obtained difference results are mapped to pixel noise intensity, thereby generating a label noise map that corresponds one-to-one with the sample image in terms of spatial resolution.

[0103] In one embodiment, since it is difficult to manually construct pairing data between noise maps and labeled noise maps, the computer device employs an online two-stage degradation method similar to Real-ESRGAN during the training phase for high-resolution images. Multiple degradation operations are dynamically added to synthesize training data that simulates real-world scenarios.

[0104] Specifically, computer equipment first processes high-definition images. The first stage of degradation processing is performed to obtain intermediate sample images. The processing procedure is as follows:

[0105]

[0106] Then at Based on this, the first degradation process in the second stage is further performed to obtain the sample image. The processing procedure is as follows:

[0107]

[0108] The symbol "→" indicates the sequential execution of image degradation processes. These degradation operations include sharpening, blurring, resolution compression, random noise superposition, compression coding, or video compression, and the intensity parameters of each operation can be dynamically and randomly generated to simulate the complex and varied noise conditions in real-world scenes as closely as possible.

[0109] To obtain the labeled data required for training, the computer device further employs a controlled variable method to introduce a degenerate reference image. The process of obtaining the degraded reference image strictly follows the same procedures as, except that it does not include noise addition. The acquisition process follows the same degradation steps and parameters as the previous one, and the specific process can be represented as follows:

[0110]

[0111]

[0112] By comparing sample images Reference diagram of degradation This generates pixel-level noise label maps, which can be used to supervise the training of noise prediction models.

[0113] In the above embodiments, the computer device acquires a high-definition original image, performs a first degradation process on the original image to obtain a sample image, and performs a second degradation process on the original image to obtain a degradation reference image. The first degradation process includes the second degradation process and noise addition process. Based on the degradation reference image and the sample image, a label noise map is determined, thereby automatically constructing pixel-level noise supervision information without manual annotation, effectively improving the efficiency of training data acquisition, ensuring the objectivity and consistency of labels, and providing a reliable guarantee for high-precision training of the noise prediction model.

[0114] In one embodiment, the process of a computer device determining a label noise map based on a degraded reference map and a sample image includes the following steps: determining a first similarity map and a second similarity map based on the degree of similarity between the degraded reference map and the sample image; fusing the first similarity map and the second similarity map to obtain a fused similarity map; and determining the label noise map based on the fused similarity distribution map.

[0115] The first similarity map refers to the pixel-level similarity distribution map obtained by comparing the degraded reference image and the sample image at the structural feature level. The structural feature comparison can be calculated based on the structural similarity index (SSIM) to reflect the degree of consistency between the two images in terms of brightness, contrast and local structure.

[0116] The second similarity map is a pixel-level similarity distribution map obtained by comparing the degraded reference image and the sample image at the level of perceptual features. Perceptual feature comparison can be calculated based on perceptual quality indicators (such as HaarPSI) to reflect the degree of visual similarity between two images.

[0117] Specifically, after acquiring the degraded reference image and the sample image, the computer device performs a pixel-level structural feature comparison between the two images. Based on structural similarity (SSIM), a first similarity map is calculated to reflect the consistency of the two images in terms of brightness, contrast, and local structure. Simultaneously, the computer device compares the two images at the perceptual feature level and calculates a second similarity map based on a perceptual quality index (such as HaarPSI) to reflect the degree of visual perceptual similarity between the two images. The first and second similarity maps are then fused, for example, by using a weighted average or non-linear weighting method, to combine the structural feature similarity and perceptual feature similarity to obtain a fused similarity map. Based on the similarity distribution in the fused similarity map, the noise intensity value of each pixel is determined, thereby generating a label noise map that corresponds one-to-one with the sample image in terms of spatial resolution.

[0118] In one embodiment, the process by which a computer device determines a labeled noise map based on a degraded reference map and a sample image can be represented as follows:

[0119]

[0120] in, Represents a degraded reference image. Represents a sample image; This represents a structural similarity index based on brightness, contrast, and structural features. This represents a human eye-perceived quality index based on wavelet transform. and In this embodiment, the fusion weight is set to 0.5 for all values ​​to balance the influence of the two similarity measures. This weighted fusion method simultaneously considers structural consistency and visual perception consistency, thereby generating a label noise map that better reflects the actual noise distribution.

[0121] In the above embodiments, the computer device determines a first similarity map and a second similarity map based on the similarity between the degraded reference map and the sample image; the first similarity map and the second similarity map are fused to obtain a fused similarity map; and a label noise map is determined based on the fused similarity distribution map. This allows for the comprehensive utilization of structural feature similarity and perceptual feature similarity, avoiding bias caused by a single indicator, improving the accuracy and stability of the generated label noise map, and providing more reliable pixel-level label data for supervised training of the noise prediction model.

[0122] In one embodiment, the process of a computer device optimizing the parameters of an initial noise prediction model based on sample noise maps and labeled noise maps to obtain a trained noise prediction model includes the following steps: determining the training loss value based on the sample noise map and labeled noise map; optimizing the parameters of the encoder and decoder of the initial noise prediction model based on the training loss value until the convergence condition is met to obtain a trained noise prediction model.

[0123] The training loss value refers to the difference measure obtained by comparing the sample noise map and the label noise map at the pixel level, and is used to characterize the degree of deviation between the prediction result and the true label.

[0124] The convergence condition refers to the condition during training where the training loss value remains within a preset threshold range and no longer decreases significantly over several consecutive iterations, or when a preset maximum number of iterations is reached. Once the convergence condition is met, the resulting noise prediction model is considered a trained model and can be used for noise prediction of target images.

[0125] Specifically, after acquiring the sample noise map and the labeled noise map, the computer device compares them at the pixel level to determine the training loss value between the sample noise map and the labeled noise map. This training loss value is used to optimize the parameters of the encoder and decoder of the initial noise prediction model. The backpropagation algorithm is used to update the network weights. During iterative training, when the training loss value decreases below a preset threshold for several consecutive rounds, or reaches a preset maximum number of iterations, the convergence condition is considered met, and the training process ends. Finally, the model obtained by the computer device is the trained noise prediction model, which can be used to predict noise in target images.

[0126] In one instance, the training loss value can be determined in the following way:

[0127]

[0128]

[0129] Where 𝑛 represents the total number of pixels. Represents the label noise map The noise intensity of the nth pixel. Represents the sample noise map The predicted noise intensity of the nth pixel is used. The L1 loss value is used to measure the difference between the prediction result and the label data at the pixel level. By minimizing this loss value, the noise prediction model can continuously optimize the parameters during training, so that the predicted noise map gradually approaches the label noise map, thereby improving the noise prediction accuracy of the model.

[0130] In the above embodiments, the computer device determines the training loss value based on the sample noise map and the label noise map; based on the training loss value, the encoder and decoder of the initial noise prediction model are optimized until the convergence condition is met, and a trained noise prediction model is obtained. This enables the accurate learning of the noise distribution characteristics at the pixel level, making the prediction results closer to the real labels, effectively improving the convergence speed and training stability of the model, and enhancing the model's generalization ability in different noise scenarios.

[0131] In one embodiment, the above-mentioned image noise level determination further includes the following steps: performing semantic segmentation on the target image to obtain a segmentation semantic map; and locating the noisy regions in the target image based on the segmentation semantic map and the adjusted noise map to obtain a noise localization map.

[0132] Semantic segmentation refers to using an image segmentation model to determine the semantic category of a target image pixel by pixel, classifying each pixel in the image into a predefined semantic category to achieve a structured understanding of the image content. In the embodiments of this application, semantic segmentation can be implemented based on a deep learning model, such as a convolutional neural network (CNN), a Transformer network, or a lightweight segmentation model (such as FastSAM), to ensure that pixel-level segmentation can be completed quickly and accurately in different scenarios.

[0133] A segmentation semantic map is a labeled image generated by the semantic segmentation process that corresponds one-to-one with the target image in terms of spatial resolution. The value of each pixel is used to characterize the semantic category to which the pixel belongs. By segmenting the semantic map, semantic boundary information of the target image at the region or object level can be obtained, providing a basis for the subsequent accurate localization of noise regions based on the adjusted noise map.

[0134] Specifically, after acquiring the target image, the computer device performs semantic segmentation on the target image to obtain a segmentation semantic map that corresponds one-to-one with the target image in terms of spatial resolution. Each pixel in the segmentation semantic map is used to represent the semantic category to which the corresponding pixel in the target image belongs. The computer device combines the segmentation semantic map with the adjusted noise map for analysis, performs statistical analysis and threshold judgment on the corresponding noise intensity in different semantic regions, determines the set of pixels in the target image whose noise exceeds a preset threshold, and generates a noise localization map based on the pixel set. The noise localization map is used to intuitively represent the areas in the target image where noise exists. Its boundary shape can be irregular and carries certain semantic information, so as to better match the human eye's visual perception of noise distribution.

[0135] For example, for a target image I, which contains three main semantic regions: "pedestrians," "vehicles," and "background buildings," FastSAM is used to perform semantic segmentation on I, resulting in a segmented semantic map. Pixels in the pedestrian area are labeled as category 1, pixels in the vehicle area are labeled as category 2, and pixels in the background building area are labeled as category 3. The corresponding adjusted noise map... The value of each pixel in the image represents the noise significance of that pixel (the higher the value, the stronger the noise). In the pedestrian area, statistical analysis of the noise intensity distribution revealed that most pixels had a noise intensity below 0.2, with only some areas reaching 0.4. In the vehicle area, the average noise intensity was 0.35, with some edge locations reaching 0.6. In the background building area, the overall noise intensity was low, with an average of around 0.1. Based on the statistical results, the system set a threshold of 0.3 for the pedestrian area, 0.4 for the vehicle area, and 0.2 for the background building area. Therefore, pixels with noise intensity > 0.3 in the pedestrian area were marked as noise areas, edge pixels with noise intensity > 0.4 in the vehicle area were marked as noise areas, and pixels with noise intensity > 0.2 in the background area were also marked, resulting in the final noise localization map. The noisy regions in the three semantic regions are shown, and these regions are consistent with the semantic information in the original image.

[0136] In the above embodiments, the computer device performs semantic segmentation on the target image to obtain a segmented semantic map; based on the segmented semantic map and the adjusted noise map, it locates the noisy regions in the target image to obtain a noise localization map. This enables the precise determination of the spatial distribution of noise at the semantic level, making the noise localization results more consistent with the visual perception of the human eye and avoiding misjudgments caused by relying solely on pixel intensity differences. At the same time, this method can output irregularly shaped noise regions that carry semantic information, providing a more reliable basis for subsequent image denoising, quality detection, and regional restoration.

[0137] In one embodiment, such as Figure 4 As shown, an image noise level determination method is also provided, which can be applied to... Figure 1 Taking a computer device (terminal 102 or server 104) as an example, the method may include the following steps:

[0138] S402, acquire a high-resolution original image; perform a first degradation process on the original image to obtain a sample image; perform a second degradation process on the original image to obtain a degradation reference image; the first degradation process includes the second degradation process and noise addition processing; determine a label noise map based on the degradation reference image and the sample image.

[0139] In one embodiment, the step of a computer device determining a label noise map based on a degraded reference map and a sample image includes the following processes: determining a first similarity map and a second similarity map based on the degree of similarity between the degraded reference map and the sample image; fusing the first similarity map and the second similarity map to obtain a fused similarity map; and determining the label noise map based on the fused similarity distribution map.

[0140] S404 extracts features from the sample image through the encoder of the initial noise prediction model to obtain multi-scale sample features and sample semantic features.

[0141] S406: The decoder of the initial noise prediction model fuses multi-scale sample features and sample semantic features to obtain sample fusion features; based on the sample fusion features, noise prediction is performed to obtain the sample noise map of the sample image.

[0142] S408: Determine the training loss value based on the sample noise map and the label noise map; optimize the parameters of the encoder and decoder of the initial noise prediction model based on the training loss value until the convergence condition is met, and obtain the trained noise prediction model.

[0143] S410 extracts features from the target image through the encoder of the noise prediction model, obtaining multi-scale features and semantic features.

[0144] S412, the multi-scale features and semantic features are fused by the decoder of the noise prediction model to obtain fused features; noise prediction is performed based on the fused features to obtain the noise map corresponding to the target image.

[0145] S414, based on the signal quantity corresponding to each pixel in the target image, adjust the noise intensity of the corresponding pixel in the noise map to obtain the adjusted noise map.

[0146] S416, Based on the adjusted noise intensity corresponding to each pixel in the target image described by the adjusted noise map, determine the noise level corresponding to each pixel.

[0147] S418, perform semantic segmentation on the target image to obtain a segmentation semantic map; based on the segmentation semantic map and the adjusted noise map, locate the noisy regions in the target image to obtain a noise localization map.

[0148] This application also provides an application scenario in which the above-described image noise level determination method is applied, with reference to... Figure 5The flowchart illustrates that the computer first performs preprocessing on the input target image, including normalization and size adjustment, brightness correction, etc., to ensure the uniformity of the input data. Then, the processed target image is input into a pre-trained noise prediction model, which employs an encoder-decoder structure: the encoder performs layer-by-layer convolution and downsampling on the image to extract edges, textures, and multi-scale features; the decoder achieves feature fusion and spatial resolution restoration through skip connections and dilated convolutions, ultimately outputting a noise map that spatially corresponds to the target image, describing the noise intensity of each pixel. During training, the computer performs multi-stage degradation processing on the high-definition original image using a degradation pool, including blurring, sharpening, noise addition, JPEG compression, resolution scaling, and video compression, to simulate complex degradation situations in real-world scenes. Based on the difference calculation between the degraded image and the reference image, a pixel-level labeled noise map is obtained, and L1Loss is used as the loss function to guide the model to continuously optimize parameters, improving prediction accuracy. The noise map gradually approximates the real label. In the application stage, after obtaining the noise map, the computer device combines the brightness and texture features of the original image to calculate the signal quantity of each pixel, and adjusts the noise intensity based on the signal-to-noise ratio formula to obtain an adjusted noise map, which is more in line with human visual perception. To further realize the spatial localization of noise areas, the computer device also introduces a FastSAM segmentation module to perform semantic segmentation on the original image to obtain a segmented semantic map. Subsequently, the computer device performs statistical joint analysis on the segmented semantic map and the adjusted noise map, and generates a noise localization map through a dynamic threshold strategy. This noise localization map can display the noise-containing areas in the target image in an irregular shape and carry certain semantic information. The final result is output by the statistical analysis module, which can intuitively reflect the intensity distribution and spatial location of the noise. Through the above process, this application can not only quantitatively analyze the image noise level in application scenarios, but also accurately locate noise areas at the semantic level, providing reliable support for image quality detection, regional noise reduction processing, and downstream visual task optimization.

[0149] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0150] Based on the same inventive concept, this application also provides an image noise level determination apparatus for implementing the image noise level determination method described above. The solution provided by this apparatus is similar to the implementation described in the above method; therefore, the specific limitations in one or more image noise level determination apparatus embodiments provided below can be found in the limitations of the image noise level determination method described above, and will not be repeated here.

[0151] In one exemplary embodiment, such as Figure 6 As shown, an image noise level determination device is provided, comprising: an image feature extraction module 602, a noise map acquisition module 604, a noise map adjustment module 606, and a noise level determination module 608, wherein:

[0152] The image feature extraction module 602 is used to extract features from the target image to obtain the image features of the target image;

[0153] The noise map acquisition module 604 is used to predict noise based on image features to obtain a noise map corresponding to the target image; the noise map describes the noise intensity corresponding to each pixel in the target image.

[0154] The noise map adjustment module 606 is used to adjust the noise intensity of corresponding pixels in the noise map based on the signal quantity corresponding to each pixel in the target image, so as to obtain the adjusted noise map.

[0155] The noise level determination module 608 is used to determine the noise level corresponding to each pixel based on the adjusted noise intensity corresponding to each pixel in the target image described by the adjusted noise map.

[0156] By extracting features from the target image, image features are obtained. Based on these features, noise prediction is performed to obtain a noise map corresponding to the target image. The noise map describes the noise intensity of each pixel in the target image, comprehensively reflecting the noise distribution of each pixel. Based on the semaphore corresponding to each pixel in the target image, the noise intensity of the corresponding pixel in the noise map is adjusted to obtain an adjusted noise map. Introducing semaphore to adjust the noise intensity allows the noise prediction results to be corrected by combining local brightness and texture characteristics, making it more consistent with the subjective perception of noise by the human eye. The resulting adjusted noise map not only improves the accuracy and robustness of noise estimation but also effectively distinguishes the saliency of noise in different regions. Based on the adjusted noise intensity of each pixel in the target image described by the adjusted noise map, the noise level of each pixel is determined, achieving fine-grained quantitative classification of image noise, thereby improving the accuracy of image noise assessment.

[0157] In one embodiment, the image features include multi-scale features and semantic features; the image feature extraction module 602 is further configured to: extract features from the target image through the encoder of the noise prediction model to obtain multi-scale features and semantic features; the noise map acquisition module 604 is configured to: fuse the multi-scale features and semantic features through the decoder of the noise prediction model to obtain fused features; and perform noise prediction based on the fused features to obtain the noise map corresponding to the target image.

[0158] In one embodiment, the apparatus further includes a model training module, configured to: acquire a sample image and a corresponding label noise map; extract features from the sample image using the encoder of an initial noise prediction model to obtain multi-scale sample features and sample semantic features; fuse the multi-scale sample features and sample semantic features using the decoder of the initial noise prediction model to obtain sample fusion features; perform noise prediction based on the sample fusion features to obtain a sample noise map of the sample image; and optimize the parameters of the initial noise prediction model based on the sample noise map and the label noise map to obtain a trained noise prediction model.

[0159] In one embodiment, the model training module is further configured to: acquire a high-resolution original image; perform a first degradation processing on the original image to obtain a sample image; perform a second degradation processing on the original image to obtain a degradation reference image; the first degradation processing includes the second degradation processing and noise addition processing; and determine a label noise map based on the degradation reference image and the sample image.

[0160] In one embodiment, the model training module is further configured to: determine a first similarity map and a second similarity map based on the similarity between the degraded reference map and the sample image; fuse the first similarity map and the second similarity map to obtain a fused similarity map; and determine a label noise map based on the fused similarity distribution map.

[0161] In one embodiment, the model training module is further configured to: determine the training loss value based on the sample noise map and the label noise map; optimize the parameters of the encoder and decoder of the initial noise prediction model based on the training loss value until the convergence condition is met, thereby obtaining the trained noise prediction model.

[0162] In one embodiment, the noise level determination module 608 is further configured to: perform semantic segmentation on the target image to obtain a segmentation semantic map; and locate the noise-containing regions in the target image based on the segmentation semantic map and the adjusted noise map to obtain a noise localization map.

[0163] Each module in the aforementioned image noise level determination device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0164] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 7 As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When executed by the processor, the computer program implements a method for determining image noise levels. The display unit is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0165] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0166] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0167] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0168] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0169] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0170] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0171] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0172] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for determining image noise level, characterized in that, The method includes: Feature extraction is performed on the target image to obtain its image features; Based on the image features, noise prediction is performed to obtain a noise map corresponding to the target image; the noise map describes the noise intensity corresponding to each pixel in the target image. Based on the signal quantity corresponding to each pixel in the target image, the noise intensity of the corresponding pixel in the noise map is adjusted to obtain the adjusted noise map; Based on the adjusted noise intensity corresponding to each pixel in the target image described by the adjusted noise map, the noise level corresponding to each pixel is determined.

2. The method according to claim 1, characterized in that, The image features include multi-scale features and semantic features; The step of extracting features from the target image to obtain image features of the target image includes: The target image is feature extracted by the encoder of the noise prediction model to obtain multi-scale features and semantic features; The step of performing noise prediction based on the image features to obtain a noise map corresponding to the target image includes: The multi-scale features and semantic features are fused using the decoder of the noise prediction model to obtain fused features; Based on the fusion features, noise prediction is performed to obtain the noise map corresponding to the target image.

3. The method according to claim 2, characterized in that, The method further includes: Obtain the sample image and the corresponding label noise map of the sample image; The encoder of the initial noise prediction model extracts features from the sample images to obtain multi-scale sample features and sample semantic features; The multi-scale sample features and the sample semantic features are fused by the decoder of the initial noise prediction model to obtain the sample fusion features; Based on the sample fusion features, noise prediction is performed to obtain a sample noise map of the sample image; Based on the sample noise map and the label noise map, the parameters of the initial noise prediction model are optimized to obtain a trained noise prediction model.

4. The method according to claim 3, characterized in that, The acquisition of the sample image and the corresponding label noise map includes: Obtain high-resolution raw images; The original image is subjected to a first degradation process to obtain a sample image; The original image is subjected to a second degradation process to obtain a degraded reference image; the first degradation process includes the second degradation process and noise addition processing. A label noise map is determined based on the degraded reference map and the sample image.

5. The method according to claim 4, characterized in that, The step of determining the label noise map based on the degraded reference map and the sample image includes: Based on the degree of similarity between the degraded reference image and the sample image, a first similarity map and a second similarity map are determined; The first similarity map and the second similarity map are merged to obtain a merged similarity map; The label noise map is determined based on the fused similarity distribution map.

6. The method according to claim 3, characterized in that, The step of optimizing the parameters of the initial noise prediction model based on the sample noise map and the label noise map to obtain a trained noise prediction model includes: The training loss value is determined based on the sample noise map and the label noise map; Based on the training loss value, the encoder and decoder of the initial noise prediction model are optimized until the convergence condition is met, thus obtaining a trained noise prediction model.

7. The method according to any one of claims 1 to 6, characterized in that, The method further includes: The target image is semantically segmented to obtain a segmented semantic map; Based on the segmentation semantic map and the adjusted noise map, the regions containing noise in the target image are located to obtain a noise localization map.

8. An image noise level determination device, characterized in that, The device includes: The image feature extraction module is used to extract features from the target image to obtain the image features of the target image; The noise map acquisition module is used to predict noise based on the image features to obtain a noise map corresponding to the target image; the noise map describes the noise intensity corresponding to each pixel in the target image. The noise map adjustment module is used to adjust the noise intensity of corresponding pixels in the noise map based on the signal quantity corresponding to each pixel in the target image, so as to obtain the adjusted noise map. The noise level determination module is used to determine the noise level corresponding to each pixel based on the adjusted noise intensity corresponding to each pixel in the target image described by the adjusted noise map.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images