Image noise evaluation method, electronic device, storage medium, and program product

Abstract

This application discloses an image noise assessment method, electronic device, storage medium, and program product, relating to the field of image quality assessment technology. It includes: performing semantic segmentation processing on the image to be assessed to obtain a first region and a second region in the image; the first region is a fine-grained entity region that interferes with actual noise recognition, and the second region is the region other than the first region; performing noise assessment on a target region of the image to be assessed to obtain a noise assessment score for the image; the target region includes at least the second region. This application can improve the accuracy of image noise assessment.

Description

Technical Field

[0001] This application relates to the field of image quality assessment technology, and in particular to an image noise assessment method, electronic device, storage medium, and program product. Background Technology

[0002] In the field of image quality assessment (IQA), image noise assessment is one of its core components. The accuracy of image noise assessment directly affects the overall quality of image processing and is of great significance for verifying the effectiveness of subsequent image processing tasks.

[0003] Currently, traditional image noise assessment methods often employ deep learning techniques, attempting to achieve noise assessment based on image noise assessment models.

[0004] However, traditional methods tend to misclassify fine-grained entities such as ground sand and gravel, and the edges of leaves as noise, resulting in poor accuracy in image noise assessment. Summary of the Invention

[0005] This application provides an image noise assessment method, electronic device, storage medium, and program product, which can at least improve the accuracy of image noise assessment.

[0006] A first aspect of this application provides an image noise assessment method, comprising: performing semantic segmentation processing on an image to be assessed to obtain a first region and a second region in the image to be assessed; the first region is a fine-grained entity region that interferes with the identification of actual noise, and the second region is a region other than the first region; performing noise assessment on a target region of the image to be assessed to obtain a noise assessment score of the image to be assessed; the target region includes at least the second region.

[0007] A second aspect of this application provides an electronic device, comprising: at least one processor; at least one memory for storing at least one program; and, when the at least one program is executed by the at least one processor, implementing the image noise evaluation method provided in any of the above aspects.

[0008] A third aspect of the embodiments of this application provides a computer-readable storage medium storing computer-executable instructions for performing the image noise evaluation method provided in any of the above aspects.

[0009] A fourth aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the image noise evaluation method provided in any of the above aspects.

[0010] The image noise assessment method provided in this application first performs semantic segmentation on the image to be assessed. Semantic segmentation divides the image into regions with different semantic meanings based on the semantic information of different objects in the image. This allows for the accurate identification of fine-grained entity regions in the image that interfere with actual noise recognition, defining them as the first region, and defining the regions other than the first region as the second region. This segmentation provides a clear and accurate target for subsequent noise assessment. After semantic segmentation, noise assessment is performed on the target regions to obtain the noise assessment score of the image. The target regions include at least the second region. Since the second region excludes fine-grained entity regions that are easily misclassified as noise, performing noise assessment on the second region avoids interference caused by misclassification in traditional methods. In summary, the image noise assessment method proposed in this application accurately distinguishes between interfering and normal regions through semantic segmentation, effectively reducing the assessment bias caused by misclassification of fine-grained entities as noise in traditional methods, thereby significantly improving the accuracy of image noise assessment. Attached Figure Description

[0011] The accompanying drawings are used to provide a further understanding of the technical solutions of this application and constitute a part of the specification. They are used together with the embodiments of this application to explain the technical solutions of this application and do not constitute a limitation on the technical solutions of this application.

[0012] Figure 1 This is a schematic flowchart of an image noise assessment method provided in an embodiment of this application; Figure 2 This is a schematic flowchart of a first noise evaluation score provided in an embodiment of this application; Figure 3 This is a schematic diagram of a second noise evaluation score provided in an embodiment of this application; Figure 4 This is a schematic diagram of the system architecture of an image noise assessment method provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages 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.

[0014] It should be understood that in the description of the embodiments of this application, the use of terms such as "first" and "second" is only for the purpose of distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of technical features indicated, or implicitly indicating the order of the technical features indicated. "At least one" refers to one or more, and "more" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any group of these items, including any group of singular or plural items. For example, at least one of a, b, and c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c can be single or multiple.

[0015] To facilitate understanding of the solutions in the embodiments of this application, some contents involved in the embodiments of this application are described below: In traditional image quality assessment techniques, the image noise assessment stage suffers from misclassification due to the similarity between fine-grained entities and real noise features. This problem stems from the inability of existing deep learning models to effectively distinguish between entities with complex textures and noise signals in images. Specifically, natural texture features such as sand and gravel surface particles and vegetation edges are incorrectly identified as noise components. Such misclassifications directly cause the noise assessment results to deviate from the true state, thereby affecting the benchmark judgment of subsequent image processing and reducing the reliability of the overall image processing workflow.

[0016] If the aforementioned problems are not addressed, inaccurate image noise assessment will lead to systemic technical consequences. Deviations in noise assessment results will propagate to the image processing stage, causing processing operations to be performed based on incorrect noise distribution information. Such processing operations not only fail to effectively suppress real noise but also damage fine-grained entity structures within the image that possess practical informational value. This chain reaction will reduce the overall output quality of the image processing system, depriving subsequent analysis tasks based on the image of a reliable data foundation, ultimately affecting the technical credibility of the image quality assessment system.

[0017] Based on this, embodiments of this application provide an image noise assessment method, electronic device, storage medium, and program product, which can improve the accuracy of image noise assessment.

[0018] The image noise assessment method provided in this application is applicable to image quality assessment tasks in complex scenarios. In many practical applications, such as satellite remote sensing image analysis, medical image diagnosis, and high-precision industrial inspection, the processed images often contain rich and complex details. These images may contain not only real noise but also a large number of fine-grained entities such as ground sand and gravel, leaf edges, and cell texture edges. The features of these fine-grained entities are similar to those of noise, which seriously interferes with the identification of actual noise. Traditional image noise assessment methods often adopt a relatively simple global assessment strategy when facing such complex images, which is difficult to effectively distinguish between fine-grained entities and real noise. It is easy to misclassify fine-grained entities as noise, resulting in inaccurate assessment results, which in turn affects the effect of subsequent image processing and may even lead to decision-making errors due to incorrect assessment. When the image noise assessment method of this application is applied to a specific scenario, it first performs semantic segmentation processing on the image to be evaluated, accurately dividing the fine-grained entity region (first region) that interferes with the identification of actual noise and the region other than the first region (second region). Subsequently, noise assessment is performed on the target region that at least contains the second region. Since the second region excludes fine-grained entity regions that are easily misclassified, it effectively avoids misclassification interference in traditional methods. In this way, a reliable noise assessment score can be obtained, providing an accurate basis for subsequent image processing and improving the accuracy and reliability of image quality assessment in complex scenes.

[0019] The image noise assessment method provided in this application will be described in detail below with reference to the accompanying drawings and through some embodiments and application scenarios.

[0020] like Figure 1 As shown, this application provides a flowchart of an image noise evaluation method, which is applied to an electronic device and includes the following steps S110 to S120: S110, perform semantic segmentation processing on the image to be evaluated to obtain a first region and a second region in the image to be evaluated; the first region is a fine-grained entity region that interferes with the recognition of actual noise, and the second region is the region other than the first region; S120, perform noise evaluation on the target region of the image to be evaluated to obtain a noise evaluation score for the image to be evaluated; the target region includes at least the second region.

[0021] For ease of understanding, the following explains some key terms in this embodiment: The image to be evaluated refers to the raw image data that requires noise level analysis and assessment. This image can be in any digital image format, such as photographs, video frames, or medical images.

[0022] Semantic segmentation is an image processing technique that aims to assign each pixel in an image to a predefined category label, thereby segmenting the image into regions with different semantic meanings. For example, it can distinguish different objects or background regions in an image, such as the sky, ground, people, and vehicles.

[0023] The first region refers to the fine-grained entity regions in the image to be evaluated, identified through semantic segmentation, that may interfere with the recognition of actual noise. These regions typically contain rich texture details, such as sand and gravel on the ground, the edges of leaves, and fabric textures, which are visually easily confused with noise.

[0024] The second region refers to all regions in the image to be evaluated other than the first region. This region typically contains relatively flat background areas with little texture or clear semantics, whose noise features are relatively easy to identify and evaluate.

[0025] The target region refers to the image region actually used to calculate the noise assessment score during noise assessment. According to this embodiment, the target region includes at least the second region, and may also include part or all of the first region.

[0026] Noise assessment refers to the process of quantitatively analyzing random or non-random interference signals present in an image in order to determine the noise level or the degree of quality degradation of the image.

[0027] A noise assessment score is a quantitative result obtained through a noise assessment process, used to characterize the noise level of an image being evaluated. This score can be a numerical value, the magnitude of which reflects the severity of the image noise.

[0028] This embodiment provides an image noise assessment method, which includes performing semantic segmentation processing on the image to be assessed to obtain a first region and a second region in the image to be assessed; and performing noise assessment on the target region of the image to be assessed to obtain a noise assessment score for the image to be assessed.

[0029] Specifically, the first step is to perform semantic segmentation on the image to be evaluated to identify different semantic regions within the image. This semantic segmentation can employ methods based on traditional image processing algorithms, such as texture analysis, color clustering, or edge detection, to divide image pixels into different categories. For example, some texture feature thresholds can be preset, marking regions with texture complexity higher than a certain threshold as fine-grained entity regions, and the remaining regions as non-fine-grained entity regions. This yields the first and second regions in the image to be evaluated. The first region is defined as a fine-grained entity region that interferes with actual noise recognition, while the second region is all regions other than the first region. Alternatively, an image segmentation method based on predefined rules can be used. For example, rules can be manually or semi-automatically defined based on the color, shape, or spatial location information of specific objects in the image to identify specific regions as the first region, while the remaining parts are automatically classified as the second region. For instance, in outdoor scene images, areas such as sand, grass, and leaves on the ground can be identified as the first region using color and texture rules.

[0030] After obtaining the first and second regions, noise evaluation is performed on the target region of the image to be evaluated to obtain a noise evaluation score for the image. Specifically, the second region of the image to be evaluated can be selected as the target region, and a certain image quality index can be directly calculated on the second region. Thus, the noise evaluation score of the image to be evaluated can be obtained directly.

[0031] The following example will provide a more detailed explanation of the above technical solution: Suppose user A took an image containing a complex outdoor scene, including ground sand and gravel, leaves, sky, and a building. This image needs a noise assessment to determine its noise level.

[0032] First, semantic segmentation is performed on the image to be evaluated. In this example, a pre-trained semantic segmentation model can be used, which can identify and distinguish different semantic categories in the image. Using this model, areas such as the ground, sand, and leaf edges are identified as regions with rich texture and fine-grained features, defined as the first region. Areas with relatively flat surfaces and less texture, such as the sky and building walls, are identified as the second region. Thus, the image to be evaluated is accurately divided into the first and second regions.

[0033] Subsequently, noise assessment is performed on the target region of the image to be evaluated. In this example, to avoid interference from fine-grained solid regions in the noise assessment, the second region is used as the primary target region for evaluation. Specifically, local statistical analysis can be performed on the second region, such as calculating the variance or standard deviation of pixel values ​​within the region, to quantify its noise level. For example, in the sky region, if the pixel values ​​fluctuate significantly, it indicates a high noise level in that region. By calculating the noise characteristics of the second region, an objective noise assessment score can be obtained.

[0034] Therefore, this method first distinguishes easily confused fine-grained entity regions (first region) from non-fine-grained entity regions (second region) in the image, and then primarily performs noise evaluation on the non-fine-grained entity regions (second region), effectively avoiding the problem of misclassifying fine-grained entities as noise in traditional methods. This combination of semantic segmentation and target region noise evaluation makes the noise evaluation results more accurate and reliable. The identification of the first region ensures the specificity of the evaluation, while the evaluation of the second region provides purer noise information. Together, they solve the technical problem of poor noise evaluation accuracy in traditional methods.

[0035] Based on the above examples, the technical concept of this embodiment demonstrates a significant technical contribution. In traditional image noise assessment methods, noise assessment is usually performed directly on the entire image. This leads to the easy misjudgment of texture details in fine-grained solid areas such as sand and gravel on the ground and the edges of leaves as noise, thereby reducing the accuracy of the assessment.

[0036] In contrast, this embodiment first performs semantic segmentation on the image to be evaluated, clearly dividing the image into a first region (fine-grained entity region) and a second region (non-fine-grained entity region). For example, in the evaluation of the outdoor scene image mentioned above, the ground sand and leaves are accurately identified as the first region, while the sky and buildings are identified as the second region. This preprocessing step is one of the key innovations of this embodiment.

[0037] Furthermore, in this embodiment, the target area is limited to at least the second region when performing noise assessment. This means that the influence of the first region can be selectively excluded or reduced when calculating the noise assessment score. For example, in the example above, by primarily assessing the noise in the sky and building areas, the inherent textures of sand and leaves are avoided from being misclassified as noise. This targeted assessment strategy allows the noise assessment results to more realistically reflect the actual noise level of the image, significantly improving the accuracy of the assessment.

[0038] In this embodiment, the image to be evaluated is first subjected to semantic segmentation. Semantic segmentation divides the image into regions with different semantic meanings based on the semantic information of different objects in the image. This allows for the accurate identification of fine-grained entity regions in the image that interfere with actual noise recognition, defining them as the first region, and defining the regions other than the first region as the second region. This segmentation provides a clear and accurate target for subsequent noise evaluation. After semantic segmentation, noise evaluation is performed on the target regions to obtain the noise evaluation score of the image. The target regions include at least the second region. Since the second region excludes fine-grained entity regions that are easily misclassified as noise, performing noise evaluation on the second region avoids interference caused by misclassification in traditional methods. In summary, the image noise evaluation method proposed in this application accurately distinguishes between interfering and normal regions through semantic segmentation, effectively reducing the evaluation bias caused by misclassification of fine-grained entities as noise in traditional methods, thereby significantly improving the accuracy of image noise evaluation.

[0039] In some embodiments described above, a semantic segmentation process is proposed to obtain a first region and a second region in the image to be evaluated. The first region is a fine-grained entity region that interferes with actual noise recognition, and the second region is the region other than the first region. Subsequently, noise evaluation is performed on the target region of the image to be evaluated to obtain a noise evaluation score. The target region includes at least the second region. However, in practical applications, how to objectively and accurately quantify the noise level of the target region is a technical problem that needs to be solved. Directly evaluating the target region may make it difficult to effectively distinguish between image content and noise, thus affecting the accuracy and reliability of the noise evaluation.

[0040] In this regard, this application further proposes that S120 includes: Obtain the first and second regions in the denoised image of the image to be evaluated; The target region of the image to be evaluated is compared and analyzed with the target region of the denoised image to obtain the noise evaluation score of the image to be evaluated.

[0041] In this embodiment, the technical feature of obtaining the first and second regions in the denoised image of the image to be evaluated aims to provide a near-noise-free reference image for noise evaluation. Specifically, a denoised image can be generated by applying a denoising algorithm to the original image to be evaluated. The denoising algorithm can employ various techniques, such as filtering-based methods like Non-Local Means (NLM) denoising and Block-matching and 3D filtering (BM3D) denoising; or deep learning-based methods, such as end-to-end denoising using a Convolutional Neural Network (CNN). After obtaining the denoised image, the corresponding first and second regions can be identified on the denoised image using the same semantic segmentation model as the original image to be evaluated.

[0042] This technique, which compares the target region of the image to be evaluated with the target region of the denoised image to obtain a noise evaluation score, quantifies noise by comparing the differences between the original and denoised images. The comparative analysis can employ various image quality assessment metrics, such as Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Mean Squared Error (MSE). These metrics quantify the degree of difference between the target region in the original image and the corresponding target region in the denoised image, which largely reflects the noise level present in the original image. Through this comparison, an objective numerical value can be obtained as the noise evaluation score of the image to be evaluated.

[0043] This application's solution, based on semantic segmentation of the image to identify the first and second regions, further introduces a denoised image as a reference. First, the image to be evaluated is denoised to generate a nearly noise-free denoised image. Then, the first and second regions corresponding to the original image are identified in the denoised image. Next, a detailed comparative analysis is performed between the target region of the image to be evaluated (e.g., the second region) and the corresponding target region in the denoised image. This comparative analysis effectively quantifies the differences between the target region in the original image and the corresponding target region in the denoised image, and these differences mainly originate from noise in the original image. In this way, the denoised image serves as a "clean" benchmark, making noise evaluation no longer dependent on subjective judgment or the difficulty of obtaining truly noise-free images, thus enabling the objective calculation of the noise evaluation score of the image to be evaluated. This method, by introducing a denoised image as a reference, provides a quantitative and operable approach to noise evaluation, enabling accurate noise evaluation even with complex image content.

[0044] Through the above technical solution, this application provides a more objective and quantitative method for image noise assessment. By introducing a denoised image as a reference and comparing the target regions of the original image and the denoised image, the image content can be effectively separated from noise, thus avoiding the blindness of noise assessment when a real noise-free image is lacking. This method enables the noise assessment score to more accurately reflect the actual noise level in the image, improving the reliability and accuracy of the assessment results. It is particularly suitable for noise assessment of complex scene images, providing a solid foundation for subsequent image processing and quality control.

[0045] In some embodiments described above in this application, a comparative analysis is proposed to obtain a noise evaluation score for the image to be evaluated by comparing the target region of the image to be evaluated with the target region of the denoised image. However, the specific method of comparative analysis is not clearly defined, resulting in poor reliability of the noise evaluation score for the image to be evaluated.

[0046] In response, this application further proposes to compare and analyze the target region of the image to be evaluated with the target region of the denoised image to obtain a noise evaluation score for the image to be evaluated, including: A first noise index is calculated between the target region of the image to be evaluated and the target region of the denoised image to obtain the noise evaluation score of the image to be evaluated. The first noise index includes at least one of peak signal-to-noise ratio and structural similarity index.

[0047] In this embodiment, the calculation of the first noise index refers to the process of quantitatively evaluating the difference or similarity between two image regions (e.g., the target region of the image to be evaluated and the target region of the denoised image) using a specific mathematical model or algorithm. Its function is to convert the noise level of the image into a measurable value, thereby achieving an objective assessment of the noise.

[0048] Peak signal-to-noise ratio (PSNR) is a widely used objective metric for measuring image quality, especially suitable for evaluating image denoising effects. It measures the noise level of the denoised image relative to the image to be evaluated. It quantifies the intensity of noise by comparing the mean square error between the original image and the denoised image. Generally, a higher PSNR value indicates less noise in the image and a better denoising effect. Specifically, PSNR can be calculated using the following formula (1):

[0049] In formula (1), Used to characterize peak signal-to-noise ratio Used to characterize the mean square error between the image A to be evaluated and the denoised image B.

[0050] The structural similarity index is an indicator designed to evaluate image quality from the perspective of human visual perception. It calculates the structural similarity between the target regions of the image being evaluated and the target regions of the denoised image within a local area using a sliding window approach. It measures the structural similarity between two images. The value of the structural similarity index typically ranges from -1 to 1, where 1 indicates that the two images are identical. A higher value indicates better image quality and better alignment with human perception of image distortion. Specifically, the structural similarity index can be calculated using the following formula (2):

[0051] In formula (2), Used to characterize structural similarity indices Used to characterize the mean value of pixel A in the image to be evaluated. The mean value of pixel B in the denoised image is used to characterize the denoised image. Used to characterize the variance of pixel values ​​in the image A to be evaluated. The variance used to characterize the B pixel value of the denoised image Used to characterize the covariance between the image A to be evaluated and the denoised image B. C1 and C2 are preset constant values ​​introduced for stability.

[0052] The above technical solution concretizes the comparative analysis between the target region of the image to be evaluated and the target region of the denoised image into the calculation of a first noise index, and clarifies that this index can include at least one of peak signal-to-noise ratio and structural similarity index. This allows image noise assessment to move beyond qualitative or vague levels and accurately characterize the noise level of an image through objective, quantitative values. This quantitative assessment method significantly improves the accuracy and reliability of noise assessment, providing a solid data foundation for subsequent image processing or quality control.

[0053] In some embodiments of this application, a noise evaluation score for the image to be evaluated is obtained by calculating a first noise index on the target region of the image to be evaluated and the target region of the denoised image. However, directly using the original first noise index result as the final noise evaluation score may result in inconsistent dimensions of the score, an unintuitive numerical range, difficulty in making comparisons across images or across indices, and may also fail to accurately reflect the subjective perception of image noise by the human eye, thereby affecting the practicality and reliability of the noise evaluation results.

[0054] In this regard, this application further proposes to calculate a first noise index between the target region of the image to be evaluated and the target region of the denoised image to obtain a noise evaluation score for the image to be evaluated, including: A first noise index is calculated between the target region of the image to be evaluated and the target region of the denoised image to obtain the first noise index result. The first noise index result is mapped to the target scaling interval through the target mapping function to obtain the scaling noise index result; Based on the scaled noise index results, the noise assessment score of the image to be evaluated is determined.

[0055] In this embodiment, the first noise index result is the raw value obtained after calculating the first noise index between the target region of the image to be evaluated and the target region of the denoised image. This result is usually an objective, quantitative indicator, such as the calculated value of peak signal-to-noise ratio or structural similarity index. These raw values ​​may have different dimensions and numerical ranges.

[0056] The target mapping function is used to transform the original primary noise index result into a more comparable and interpretable value. Its function is to perform a non-linear or linear transformation on the original index to adjust its numerical range and distribution, making it more suitable for specific evaluation needs. The target scaling interval is a preset numerical range used to normalize the primary noise index result. Through the action of the target mapping function, the original primary noise index result will be mapped to this target scaling interval, thus making the noise evaluation results of different indices or different images have uniform dimensions and comparability. Common scaling intervals include [0, 10], [0, 100], etc.

[0057] The scaled noise index result is a value that falls within the target scaling interval after processing by the target mapping function. This scaled noise index result represents the performance of the original first noise index result on a standardized scale, and has better interpretability and comparability. The step of determining the noise evaluation score of the image to be evaluated refers to obtaining the final noise evaluation score of the image to be evaluated based on the scaled noise index result. This score can be the scaled noise index result itself, or it can be a final value obtained by further rounding, integerization, or combining other rules, used to intuitively represent the noise level of the image.

[0058] Specifically, the noise evaluation score of the image to be evaluated can be calculated using the following formula (3):

[0059] In formula (3), A noise evaluation score used to characterize the image being evaluated. Used to characterize the peak signal-to-noise ratio of scaling. A structural similarity index used to characterize scaling.

[0060] This application's solution calculates a first noise index between the target region of the image to be evaluated and the target region of the denoised image. Based on the first noise index result, a scaling mechanism is further introduced. Specifically, while the original first noise index results, such as PSNR or SSIM values, objectively reflect the noise level of the image, their numerical range and physical meaning may not be intuitive, making them difficult to directly compare between different scenes or different indices. To address this issue, this application applies a target mapping function to map these original first noise index results to a preset target scaling interval, thereby obtaining scaled noise index results. This process effectively unifies the dimensions and numerical ranges of different noise indices, making the evaluation results more comparable and interpretable. Finally, based on these scaled noise index results, the noise evaluation score of the image to be evaluated can be determined more accurately and intuitively. In this way, the solution of this application can provide a standardized, easy-to-understand noise evaluation score that is more consistent with human visual perception, significantly improving the practicality and reliability of the noise evaluation results.

[0061] By employing the aforementioned technical solution, the original first noise index result is mapped to the target scaling interval through a target mapping function, effectively addressing the issues of inconsistent numerical ranges and poor interpretability of the original noise index. This allows evaluation results for different images or different noise indices to be compared on a uniform scale, improving the intuitiveness and understandability of noise evaluation scores.

[0062] In some embodiments of this application described above, a first noise index is calculated between the target region of the image to be evaluated and the target region of the denoised image. After obtaining the first noise index result, it is necessary to map the first noise index result to a target scaling interval to determine the noise evaluation score of the image to be evaluated. However, if the target mapping function is not properly selected, the mapped scaling noise index result may not accurately reflect the nonlinear variation characteristics of the noise index result, thereby affecting the accuracy and robustness of the noise evaluation score.

[0063] In this regard, this application further proposes that the target mapping function is a logistic function composed of exponential functions with the natural constant as the base.

[0064] In this embodiment, the logistic function is an S-shaped curve function, characterized by its ability to map any real input value to a finite output range, such as 0 to 10 or 0 to 100. This function is typically constructed from an exponential function with the natural constant e as its base. Due to its smoothness, continuity, and nonlinear characteristics, the logistic function is widely used in fields such as statistics, machine learning, and signal processing for data normalization, probabilistic modeling, and nonlinear transformations.

[0065] As an example, when the first noise index result is the peak signal-to-noise ratio, its corresponding target mapping function can be shown in the following formula (4):

[0066] In formula (4), Used to characterize the peak signal-to-noise ratio of scaling. The peak signal-to-noise ratio is used to characterize the peak signal-to-noise ratio, and e is used to characterize the natural constant.

[0067] When the first noise index result is a structural similarity index, its corresponding target mapping function can be shown in the following formula (5):

[0068] In formula (5), Structural similarity index used to characterize scaling The index is used to characterize the structural similarity index, and e is used to characterize the natural constant.

[0069] The proposed solution applies a logistic function, constructed from an exponential function with a natural constant base, to the mapping process of the first noise index result. This effectively transforms the original first noise index result (e.g., peak signal-to-noise ratio or structural similarity index) non-linearly from its original numerical range to a preset target scaling interval. The S-shaped curve characteristic of the logistic function makes it highly sensitive to changes in the middle region of the input value, while it tends to flatten out at both ends. This means that for the first noise index results in the middle range that have a greater impact on the perception of image quality, the logistic function can provide a finer and more sensitive mapping, thus more accurately reflecting subtle quality differences. For extreme high-quality or low-quality regions, it can smoothly map the index value to the boundary of the scaling interval, avoiding oversensitivity or unreasonable jumps in the evaluation results caused by excessively high or low index values. This non-linear mapping method can better simulate the human eye's perception of image noise, because the human eye's sensitivity to noise is not non-linear, but rather exhibits a characteristic of higher sensitivity within a specific range and lower sensitivity within other ranges. In this way, the original first noise index result is effectively "compressed" or "stretched" into the target scaling range, thereby obtaining a more representative and robust scaling noise index result, and thus determining a more accurate noise assessment score.

[0070] By employing the logistic function, a function with an exponential function based on the natural constant, as the target mapping function, this application can nonlinearly transform the original first noise index results (such as peak signal-to-noise ratio or structural similarity index), making them more accurately mapped to the target scaling interval. This nonlinear mapping characteristic, especially its S-shaped curve, better simulates the human eye's perception of image noise; that is, it is more sensitive to quality changes within a specific noise level range, while tending to stabilize at extremely high or low noise levels. This avoids the problem of evaluation results not matching subjective perception that may occur with traditional linear mapping, improving the accuracy and robustness of noise evaluation scores.

[0071] In some embodiments described above in this application, noise assessment of target regions of an image is proposed, which requires obtaining a denoised image of the image to be assessed. However, in practical applications, directly denoising the entire image to be assessed may result in high computational resource consumption, low processing efficiency, or unsatisfactory denoising effects due to factors such as excessively large image size or uneven noise distribution, thereby affecting the accuracy of subsequent noise assessment.

[0072] In this regard, this application further proposes that, before obtaining the first region and the second region in the denoised image of the image to be evaluated, the method further includes: The image to be evaluated is divided into multiple image blocks; Denoising is performed on each image block separately to obtain multiple denoised image blocks; The denoised image blocks are stitched together to obtain the denoised image of the image to be evaluated.

[0073] In this embodiment, dividing the image to be evaluated into multiple image blocks means decomposing the original image to be evaluated into several smaller sub-image units. This division can be performed using a fixed-size sliding window, with overlapping areas between windows to ensure smoothness of subsequent stitching, or it can be a non-overlapping regular grid division; alternatively, it can be an adaptive non-uniform division based on image content features, such as edges, textures, or salient regions.

[0074] Denoising each image block separately refers to applying a denoising algorithm to each individual image block to remove the noise it contains. This can be done using various denoising techniques, such as deep learning-based denoising models (e.g., convolutional neural network denoising models, generative adversarial network denoising models, etc.), or traditional signal processing methods (e.g., wavelet thresholding denoising, nonlocal mean filtering, 3D block matching denoising, etc.).

[0075] Stitching together the denoised image blocks to obtain the denoised image to be evaluated refers to recombining all the image blocks that have undergone independent denoising processing to restore the complete size and structure of the original image to be evaluated. If there is overlap between image blocks, techniques such as weighted averaging, Poisson fusion, or multi-resolution stitching can be used during stitching to eliminate stitching artifacts and ensure the overall visual coherence and smoothness of the image; if there is no overlap between image blocks, they can be simply combined according to their relative positions in the original image.

[0076] The proposed solution first decomposes the image to be evaluated into multiple smaller image blocks, allowing subsequent denoising processing to target these local regions. This block-based processing approach effectively reduces the computational complexity of a single denoising operation, improving processing efficiency. Furthermore, the smaller image block size allows for more precise capture and processing of noise features in local areas, thereby enhancing the accuracy and effectiveness of denoising. After denoising each image block separately, these locally optimized denoised image blocks are seamlessly stitched together to obtain a high-quality denoised image with excellent overall denoising performance. This strategy, combining local processing with global reconstruction, effectively overcomes the problems of low computational efficiency and poor denoising results that may be encountered when directly denoising the entire image, providing a reliable foundation for accurate noise evaluation.

[0077] The above technical solution involves dividing the image to be evaluated into blocks, denoising each block separately, and then stitching them together. This effectively solves the problems of low computational efficiency and poor denoising effect that may occur when directly denoising the entire image. This processing method can handle noise in local areas of the image more precisely, improving the quality and generation efficiency of the denoised image. This lays the foundation for subsequent comparative analysis between the target area of ​​the image to be evaluated and the target area of ​​the denoised image to obtain a more accurate and reliable noise evaluation score.

[0078] In some of the embodiments described above in this application, due to the differences in characteristics of different regions in the image to be evaluated, especially the first region (a fine-grained entity region that interferes with the actual noise identification) and the second region (regions other than the first region), if a single or indiscriminate noise evaluation method is used, the noise evaluation score may not be able to fully and accurately reflect the true noise status of the image.

[0079] In this regard, this application further proposes that the noise assessment score includes at least one of a first noise assessment score and a second noise assessment score; The first noise assessment score is the score obtained after performing noise assessment on the second region of the image to be evaluated. The second noise assessment score is a weighted composite score obtained by assessing the noise in the first and second regions of the image to be evaluated separately.

[0080] In this embodiment, the noise evaluation score is a quantitative indicator measuring the noise level of an image. The first noise evaluation score is obtained by specifically evaluating the noise in the second region of the image to be evaluated. Evaluating the second region means applying the noise evaluation algorithm to a region of the image that has minimal interference with actual noise identification. The second noise evaluation score is a weighted composite score obtained by evaluating the noise in the first and second regions of the image separately. Evaluating the first and second regions separately means applying the noise evaluation algorithm independently to each region, obtaining their respective noise evaluation results. These independent evaluation results are then combined according to preset or dynamically determined weights to form a unified composite score.

[0081] This application's solution introduces diverse noise assessment score types, making the noise assessment process more flexible and targeted. When focusing only on a relatively uniform second region in the image where noise features are easier to identify, the first noise assessment score can be directly obtained, thus avoiding potential interference from fine-grained entities in the first region and improving the accuracy and reliability of the assessment. On the other hand, when a comprehensive consideration of the overall noise situation of the image is required, noise assessments can be performed on the first and second regions separately, and weighted according to their respective characteristics to obtain the second noise assessment score. This weighted comprehensive assessment method not only takes into account the noise contribution of different regions but also effectively balances the interference of the first region through weight adjustment, making the final noise assessment score more comprehensive and objective, and better adaptable to different image content and assessment needs.

[0082] like Figure 2 The diagram illustrates a flowchart for obtaining a first noise evaluation score. Specifically, the process begins by acquiring the image to be evaluated (i.e., the image to be evaluated) and its corresponding denoised image (i.e., the denoised image). Semantic segmentation is then performed on both images to distinguish between fine-grained entity regions (i.e., the first region) that interfere with noise recognition and normal regions (i.e., the second region). Subsequently, only the normal regions (i.e., the second region) of both images are evaluated, and the peak signal-to-noise ratio (PSNR) and structural similarity index are calculated between them. Finally, based on the PSNR and structural similarity index results for the normal regions, a final noise score of 1 is determined. This noise score is the first noise evaluation score obtained by evaluating only the normal regions (i.e., the second region).

[0083] like Figure 3The diagram illustrates a flowchart for a second noise evaluation score. Specifically, first, the test image (i.e., the image to be evaluated) and its corresponding denoised image (i.e., the denoised image) are acquired. Semantic segmentation is then performed on both images to distinguish between fine-grained entity regions (i.e., the first region) that interfere with noise recognition and normal regions (i.e., the second region). Subsequently, noise indices are calculated in two paths: one path calculates the peak signal-to-noise ratio (PSNR) and structural similarity index for the normal region (i.e., the second region) and generates a noise score 1; the other path calculates the PSNR and structural similarity index for the abnormal region (i.e., the first region) and generates an abnormal region noise score. Simultaneously, based on the texture of the normal and abnormal regions in the test image (i.e., the image to be evaluated), weight values ​​for the first and second regions are determined. Finally, the noise scores for the normal and abnormal regions are weighted and fused based on these weight values ​​to output a comprehensive noise score 2, which is the second noise evaluation score that considers both regions simultaneously.

[0084] Through the above technical solution, this application can provide more flexible and refined image noise assessment results. By distinguishing between the first noise assessment score and the second noise assessment score, noise assessment results can be selectively obtained only for non-interference areas in the image, depending on different application scenarios and assessment focuses. This avoids potential interference from fine-grained physical regions on noise assessment and improves the accuracy of the assessment. Simultaneously, by performing weighted comprehensive assessment of different regions, the overall noise distribution and characteristics of the image can be fully reflected, making the assessment results more representative and robust, effectively solving the problem that a single assessment method cannot take into account the noise characteristics of different regions.

[0085] In some embodiments described above, semantic segmentation of the image to be evaluated is proposed to obtain a first region and a second region. Noise evaluation is then performed on each of these two regions, followed by weighted summation to obtain a comprehensive score. However, in practical applications, simply weighting the noise evaluation results of different regions without fully considering the characteristics of each region may result in a final noise evaluation score that fails to accurately reflect the overall noise level of the image. This is especially true when there are significant differences in the sensitivity to noise or the noise manifestation patterns of different regions, which can affect the reliability of the evaluation results.

[0086] In this regard, this application further proposes that the noise assessment score includes a second noise assessment score; and that S120 also includes: Based on the difference of a second noise index between a first region and a second region of the image to be evaluated, a first evaluation weight for the first region and a second evaluation weight for the second region are determined. The second noise metric includes at least one of texture degree and area ratio.

[0087] In this embodiment, the second noise metric is a parameter used to quantify the inherent characteristics of an image region, and its differences can serve as an objective basis for guiding weight allocation. For example, texture can characterize the complexity or detail richness of an image region, and can be quantified by calculating features such as the contrast of the gray-level co-occurrence matrix (GLCM). Area proportion refers to the proportion of pixels occupied by a certain region in the entire image, which can be obtained by calculating the ratio of the number of pixels in the region to the total number of pixels in the image.

[0088] The first evaluation weight of the first region and the second evaluation weight of the second region are determined based on the difference in a second noise index between the first and second regions of the image to be evaluated. This aims to dynamically adjust the contribution of each region to the overall noise assessment based on their unique characteristics. This can be achieved in several ways; for example, the weights can be determined based on the relative magnitude of the second noise index, or a pre-defined mapping function (such as a linear function, sigmoid function, or exponential function) can be used to map the differences in the second noise index to weight values. The first and second evaluation weights quantify the relative importance or contribution of the first and second regions to the final overall noise assessment score. They are typically values ​​between 0 and 1, and the sum of the weights for all regions is 1 to ensure that the noise assessment results from different regions can be fused in a reasonable manner.

[0089] The proposed solution divides the image into a first region and a second region through semantic segmentation. Before evaluating noise in these regions and weighting them to obtain a comprehensive score, a step is introduced to determine the evaluation weights based on the differences in a second noise index. Specifically, the system calculates the second noise index (e.g., texture, area ratio) for each of the first and second regions and analyzes the differences between these indices. For example, if the first region (e.g., fine-grained entities such as faces and text) has high texture and a small area ratio, while the second region (e.g., background, sky) has low texture and a large area ratio, the system dynamically assigns the first and second evaluation weights based on these differences. This dynamic weight allocation mechanism ensures that the noise evaluation results of different regions can be fused in a more reasonable way that conforms to human visual perception. In this way, the proposed solution overcomes the limitations of simple weighting or fixed weighting, enabling the final noise evaluation score to more accurately reflect the overall noise level of the image. Especially when the image content is complex and the characteristics of each region differ significantly, the reliability and accuracy of the evaluation results are significantly improved.

[0090] As an example, the first area percentage of the first region can be determined by the following formula (6):

[0091] In formula (6), Used to characterize the first area percentage of the first region. Used to characterize the area of ​​the first region Used to characterize the area of ​​the second region.

[0092] The second area proportion of the second region can be determined by the following formula (7):

[0093] In formula (7), Used to characterize the second area proportion of the second region. Used to characterize the area of ​​the first region Used to characterize the area of ​​the second region.

[0094] After obtaining the first evaluation weight of the first region and the second evaluation weight of the second region, the second noise evaluation score can be determined by the following formula (8):

[0095] In formula (8), Used to characterize the second noise evaluation score The second evaluation weight is used to characterize the second region. The first evaluation weight is used to characterize the first region. Used to characterize the score obtained after noise assessment of the second region. This score is used to characterize the score obtained after noise assessment of the first region.

[0096] Through the above technical solution, this application can dynamically adjust the contribution of different regions in the comprehensive noise assessment based on their inherent characteristics (such as texture and area ratio) in the image to be evaluated. This allows the noise assessment score to more accurately reflect the sensitivity and manifestation of noise in different regions, thereby improving the objectivity and reliability of the comprehensive noise assessment score. By quantitatively analyzing the characteristics of different regions and dynamically adjusting the weights accordingly, this solution can more precisely capture the differences in noise perception in different parts of the image, making the final assessment result more consistent with human visual perception, thus improving the accuracy and practicality of image noise assessment.

[0097] In some embodiments described above, to achieve a more refined evaluation of image noise, it is necessary to determine the corresponding evaluation weights based on the characteristics of different regions in the image. Texture is one of the important factors affecting weight allocation. However, in practical applications, how to accurately and quantitatively characterize the texture of image regions to ensure the rationality of subsequent weight allocation and the accuracy of evaluation results is a technical problem that needs to be solved. If the texture characterization is not accurate enough or is ambiguous, it may affect the accuracy of noise evaluation weights for different regions, thereby affecting the reliability of the final noise evaluation score.

[0098] In this regard, this application further proposes that texture is characterized by the contrast of the gray-scale feature matrix.

[0099] In this embodiment, texture is an important indicator for describing the visual features of an image region. It reflects the spatial variation and repetition patterns of pixel grayscale values ​​in the image. High texture usually means that the image content is complex and rich in detail; low texture indicates that the image content is relatively smooth and uniform. Accurate characterization of texture is crucial for distinguishing the characteristics of different regions in an image, especially in noise assessment, where different texture regions may have different sensitivities to noise.

[0100] A Gray-Level Co-occurrence Matrix (GLCM) is a statistical method used to describe the distribution of gray values ​​between two pixels in an image that have a specific spatial relationship (e.g., adjacent, separated by a certain distance). It is constructed by statistically analyzing the frequency of occurrence of pixels with specific gray-level value pairs in the image. Contrast is a key feature of GLCM, reflecting the drastic degree of local gray-level value changes in an image. A higher contrast value indicates a coarser image texture and more pronounced gray-level changes; a lower contrast value indicates a finer image texture and smoother gray-level changes.

[0101] The proposed solution characterizes texture by specifically representing it through the contrast of gray-level feature matrices, making the quantification of texture characteristics of image regions more explicit and objective. Specifically, for the first and second regions in the image to be evaluated, their gray-level feature matrices are calculated, and contrast values ​​are extracted from these matrices. These contrast values ​​directly reflect the degree of gray-level variation in each region, i.e., the coarseness or fineness of its texture. In this way, quantified texture indices can be obtained, providing a reliable basis for subsequently determining the first evaluation weight of the first region and the second evaluation weight of the second region based on texture differences. This characterization method avoids subjective judgment, standardizes the texture acquisition process, ensures the objectivity and accuracy of the evaluation weights, and thus improves the reliability of the overall noise evaluation score.

[0102] As an example, the first texture degree of the first region can be determined by the following formula (9):

[0103] In formula (9), The first texture degree is used to characterize the first region, and i and j are used to characterize the pixel gray level index in the gray-level co-occurrence matrix, respectively. The square of the difference between gray level i and gray level j is the weight of the contrast. Used to characterize the position in the gray-level co-occurrence matrix The value represents the joint probability that gray level i and gray level j appear in a specific spatial relationship in the first region.

[0104] The second texture degree of the second region can be determined by the following formula (10):

[0105] In formula (10), The second texture degree is used to characterize the second region, and i and j are used to characterize the pixel gray level index in the gray-level co-occurrence matrix, respectively. The square of the difference between gray level i and gray level j is the weight of the contrast. Used to characterize the position in the gray-level co-occurrence matrix The value represents the joint probability that gray level i and gray level j appear in a specific spatial relationship in the second region.

[0106] The above technical solution characterizes texture through the contrast of the gray-level feature matrix, solving the problem of accurately quantifying the texture of image regions in image noise assessment. This quantification method provides objective and repeatable texture feature indicators, enabling more precise judgments based on texture differences when determining the evaluation weights of the first and second regions. This avoids weight allocation bias caused by inaccurate texture characterization, thereby improving the accuracy and reliability of noise assessment scores, especially when processing images containing complex texture details, enabling more effective identification and evaluation of noise levels in different regions.

[0107] In some embodiments described above in this application, a first evaluation weight for the first region and a second evaluation weight for the second region are determined based on the difference in a second noise index between a first region and a second region of the image to be evaluated. However, it does not specifically explain how the first evaluation weight for the first region and the second evaluation weight for the second region are determined.

[0108] In this regard, this application further proposes to determine a first evaluation weight for the first region and a second evaluation weight for the second region based on the difference in a second noise index between a first region and a second region of the image to be evaluated, including: Based on the first texture degree of the first region and the second texture degree of the second region of the image to be evaluated, the first texture weight of the first region and the second texture weight of the second region are determined. The first initial weight is determined based on the first texture weight and the first area ratio of the first region, and the second initial weight is determined based on the second texture weight and the second area ratio of the second region. The first initial weight and the second initial weight are normalized to obtain the first evaluation weight of the first region and the second evaluation weight of the second region.

[0109] In this embodiment, based on the first texture degree of the first region and the second texture degree of the second region of the image to be evaluated, the first texture weight of the first region and the second texture weight of the second region are determined. This step aims to quantify the relative importance of texture features in noise assessment based on the texture features of different regions of the image. Texture degree is an indicator of the visual complexity of an image region. It is generally believed that the more complex the texture of a region, the more sensitive it may be to noise perception or the more diverse the noise manifestations. By calculating the first texture degree of the first region and the second texture degree of the second region, the texture information of these two regions can be obtained respectively. Based on this, various methods can be used to determine the texture weight. For example, the texture weight can be determined based on the reciprocal of the texture degree, that is, the higher the texture degree, the smaller its reciprocal, and the lower the corresponding texture weight can be set, and vice versa, to reflect the sensitivity of complex texture regions to noise assessment. Alternatively, a preset mapping function or lookup table can be used to map the texture degree value to a weight value between 0 and 1. This mapping function can be adjusted according to experimental data or expert experience to optimize the weight allocation.

[0110] The first initial weight is determined based on the first texture weight and the first area proportion of the first region, and the second initial weight is determined based on the second texture weight and the second area proportion of the second region. This step comprehensively considers the texture complexity of the region and its spatial proportion (i.e., area proportion) in the image to form a preliminary, unnormalized weight. The area proportion of a region reflects the visual salience or information content of the region in the entire image. The larger the area, the greater the impact on the overall noise perception. By combining texture weight and area proportion, the overall importance of each region can be evaluated more comprehensively. Specifically, the first initial weight can be determined by the following formula (11):

[0111] In formula (11), The first initial weights are used to characterize the first region. Used to characterize the first area percentage of the first region. The first texture weight is used to characterize the first region.

[0112] The second initial weight can be determined by the following formula (12):

[0113] In formula (11), The second initial weights are used to characterize the second region. Used to characterize the second area proportion of the second region. The second texture weight is used to characterize the second region.

[0114] The first and second initial weights are normalized to obtain the first evaluation weight for the first region and the second evaluation weight for the second region. This step aims to ensure that the final evaluation weights are comparable and operable, allowing them to be used as weighting coefficients in subsequent noise evaluation calculations. Normalization adjusts the initial weights to a uniform scale range, for example, making the sum of all evaluation weights equal to 1. Specifically, the first evaluation weight can be determined using the following formula (13):

[0115] In formula (13), The first evaluation weight is used to characterize the first region. The first initial weights are used to characterize the first region. First initial weights used to characterize the first region With the second initial weight of the second region The sum of .

[0116] The second evaluation weight can be determined by the following formula (14):

[0117] In formula (14), The second evaluation weight is used to characterize the second region. The second initial weights are used to characterize the second region. First initial weights used to characterize the first region With the second initial weight of the second region The sum of .

[0118] This application's scheme overcomes the limitations of relying solely on single-indicator differences by structurally combining the texture complexity and spatial proportion of image regions to determine noise evaluation weights. First, for the first and second regions obtained from semantic segmentation in the image to be evaluated, their texture degrees are calculated, and their respective texture weights are determined based on these texture degrees. This process ensures the fundamental role of texture features in weight allocation, allowing regions with complex textures to receive weights commensurate with their noise sensitivity. Subsequently, these texture weights are fused with the area proportion of the corresponding regions to generate first and second initial weights. This fusion mechanism cleverly combines the visual complexity (texture) of a region with its spatial importance (area) in the image, ensuring that the weights not only reflect the region's intrinsic characteristics but also consider its contribution to the overall image. Finally, these initial weights are normalized to ensure that the final first and second evaluation weights are numerically consistent and operable, and can be directly used for weighted calculations. Through this series of steps, this application's scheme can systematically and comprehensively quantify the relative importance of different regions in noise evaluation, providing a more accurate and reasonable basis for subsequent weighted comprehensive evaluation.

[0119] Through the above technical solution, this application provides a more refined and accurate method for determining region evaluation weights. By comprehensively considering the texture complexity of a region and its spatial proportion in the image, and performing systematic weight calculation and normalization, the inaccurate weight allocation caused by relying solely on differences in a single indicator can be avoided. This allows the final noise evaluation score to more realistically reflect the noise perception characteristics of different regions of the image. Especially when the image content is complex and the noise sensitivity of different regions varies greatly, it can significantly improve the reliability and robustness of the noise evaluation results, thereby providing more effective guidance for image processing and quality control.

[0120] In some embodiments described above in this application, a first evaluation weight for the first region and a second evaluation weight for the second region are determined based on the difference in a second noise index between a first region and a second region of the image to be evaluated. The second noise index includes at least one of texture degree and area ratio. However, if texture degree is directly used as a weighting factor when determining the texture weight of each region, it may not accurately reflect the impact of texture complexity on the reliability of noise evaluation, leading to unreasonable weight allocation and consequently affecting the accuracy of the final noise evaluation score.

[0121] In this regard, this application further proposes determining the first texture weight of the first region and the second texture weight of the second region based on the first texture degree of the first region and the second texture degree of the second region of the image to be evaluated, including: The first texture weight of the first region is determined based on the reciprocal of the first texture degree. The second texture weight of the second region is determined based on the reciprocal of the second texture degree.

[0122] In this embodiment, the first texture degree and the second texture degree refer to the texture complexity or features of the first and second regions in the image to be evaluated. They are indicators that measure the gray-level changes or structural repetition of local image regions and can characterize the richness of detail in a region. The reciprocal refers to the multiplicative inverse of a number, which introduces an inverse relationship, such that the higher the texture degree, the smaller the corresponding reciprocal value. Therefore, in the subsequent weight calculation, regions with high texture degree are assigned relatively low weights, reflecting the interference or uncertainty of complex texture regions on noise evaluation. The first texture weight and the second texture weight are intermediate quantities used for subsequent calculation of the first initial weight and the second initial weight. They reflect the importance or reliability of the first and second regions in noise evaluation based on texture features. These weights will be used to adjust the contribution of each region to the final noise evaluation score, ensuring that the influence of texture features on the evaluation results is reasonably reflected.

[0123] Specifically, the first texture weight can be determined by the following formula (15):

[0124] In formula (15), The first texture weight is used to characterize the first region. The first texture degree is used to characterize the first region.

[0125] The second texture weight can be determined by the following formula (16):

[0126] In formula (16), Second texture weights used to characterize the second region The second texture degree is used to characterize the second region.

[0127] The proposed solution cleverly utilizes the reciprocal of texture degree when determining the first texture weight of the first region and the second texture weight of the second region. Specifically, instead of directly using the first texture degree, the first texture degree is used to calculate the first texture weight of the first region; instead, its reciprocal is used. Similarly, the second texture weight of the second region is also determined based on its reciprocal. The underlying logic of this approach is that, generally, the more complex the texture of an image region (i.e., the higher the texture degree), the richer its internal details. This can make it difficult to distinguish noise from texture details, thus reducing the reliability of noise evaluation in that region. By using the reciprocal of texture degree as the texture weight, regions with high texture degree can be effectively assigned relatively lower texture weights, while regions with low texture degree receive relatively higher texture weights. This inverse weighting mechanism ensures that the negative impact of texture complexity on region evaluation weights is reasonably reflected in the subsequent calculation of the first and second initial weights. Ultimately, this texture weight, adjusted by the inverse of texture degree, combined with the area proportion of each region and then normalized, yields a more robust and accurate first and second evaluation weight. This allows the final noise evaluation score to more accurately reflect the actual noise level of the image, especially in the presence of fine-grained solid regions.

[0128] The above technical solution, when determining the first texture weight of the first region and the second texture weight of the second region, no longer simply relies on the texture degree itself, but cleverly utilizes the reciprocal of the texture degree. This reciprocal-based weight determination method can ensure that the influence of high texture degree regions is appropriately reduced, thereby avoiding excessive influence of complex texture regions on the noise score.

[0129] In some of the embodiments described above in this application, semantic segmentation is performed on the image to be evaluated to obtain a first region and a second region. However, in practice, how to accurately and effectively extract and define these regions from the semantic segmentation results to ensure the accuracy of subsequent noise evaluation is a technical problem that needs to be solved.

[0130] In this regard, this application further proposes that S110 includes: Semantic segmentation is performed on the image to be evaluated to obtain the mask image corresponding to the image to be evaluated; Based on the mask image, the first and second regions in the image to be evaluated are determined.

[0131] In this embodiment, semantic segmentation is performed on the image to be evaluated to obtain a mask map corresponding to the image. The purpose is to classify each pixel in the image into a predefined semantic category, thereby identifying the boundaries and regions of different objects in the image. The obtained mask map is the direct output of semantic segmentation, representing the distribution of different semantic regions in the image with pixel-level precision, providing a foundation for subsequent region determination. This processing can employ deep learning-based semantic segmentation models, which learn image features through convolutional neural networks and output a category prediction for each pixel, thereby generating a pixel-level semantic mask. Alternatively, traditional image processing methods, such as threshold segmentation, region growing, edge detection combined with morphological operations, can be used to preprocess and extract features from the image, and then classify pixels according to preset rules or classifiers to generate the corresponding mask map.

[0132] Based on the mask image, a first region and a second region in the image to be evaluated are determined. Their function is to divide the image into two regions with different meanings for noise evaluation based on the mask image obtained from semantic segmentation: a first region (fine-grained entity regions) and a second region (regions other than the first region). This ensures that subsequent noise evaluation can be differentiated for different types of regions, improving the accuracy and robustness of the evaluation. Specifically, the first and second regions can be determined by analyzing the pixel distribution of different semantic categories in the mask image. For example, if the mask image segments the image into categories such as "face," "background," and "clothing," fine-grained entity regions such as "face" can be merged into the first region, while other regions such as "background" can be merged into the second region.

[0133] The proposed solution performs semantic segmentation on the image to be evaluated, aiming to classify each pixel in the image into a predefined semantic category, thereby generating a pixel-level mask. This mask visually represents the spatial distribution and boundary information of different semantic objects in the image. Subsequently, based on this mask, the system can clearly identify and divide a first region and a second region. Specifically, by parsing the pixel information of different semantic categories in the mask, semantic categories representing fine-grained entities that interfere with actual noise recognition (such as faces, text, complex textures, etc.) can be aggregated into the first region, while the set of pixels of all other semantic categories or those not identified as the first region is defined as the second region. This method of generating a mask through semantic segmentation and then determining the region makes the division of the first and second regions highly automated and accurate, avoiding the limitations of vague region division or manual intervention that may exist in traditional methods. In this way, subsequent noise assessments can be tailored to different regions. For example, a more refined noise assessment strategy may be needed for fine-grained physical regions (the first region), while a more general assessment method may be used for background regions (the second region), thereby improving the accuracy and reliability of the overall noise assessment.

[0134] Using the above technical solution, after semantic segmentation of the image to be evaluated, an accurate mask map can be obtained, which can then be used to accurately determine the first and second regions. This region division method based on semantic segmentation and mask maps overcomes the ambiguity and subjectivity that may exist in traditional methods, making the definition of the first region (fine-grained entity region) and the second region (non-fine-grained entity region) more automated and refined. This provides a solid foundation for subsequent differentiated noise evaluation for different regions, ensuring the accuracy and reliability of noise evaluation. Especially when processing images containing complex scenes and diverse content, it can effectively avoid interference with the noise evaluation of fine-grained entity regions, thereby improving the robustness of the overall image noise evaluation.

[0135] It should be noted that the various embodiments described in this application can be combined with each other or implemented individually without conflict, and this application does not limit this.

[0136] For ease of understanding, a specific embodiment will be used as an example: like Figure 4The diagram illustrates a system architecture for an image noise assessment method. The image to be tested (i.e., the image to be evaluated) undergoes preprocessing via two parallel branches. One branch inputs a denoising model, generating a denoised image through block denoising and subsequent stitching. The other branch inputs a semantic segmentation model to generate a mask image, accurately dividing the abnormal regions (first region) and normal regions (second region) that are prone to noise interference into the mask, simultaneously obtaining the corresponding regions of the denoised image. The preprocessing results are then fed into an integrated computation module. Computation module 1, focusing only on the normal region, calculates the peak signal-to-noise ratio and structural similarity index of the region in the image to be tested (i.e., the image to be evaluated) and the denoised image, and outputs a first noise assessment score (noise score 1) after scaling using a logistic function. Computation module 2 first determines the evaluation weights of the two regions based on the inverse of their texture (grayscale feature matrix contrast) and area ratio, then combines the noise index results of the abnormal and normal regions with weighted fusion to finally output a comprehensive second noise assessment score (noise score 2), achieving accurate noise assessment that takes into account regional characteristics.

[0137] This application proposes an image noise evaluation method combining semantic segmentation. The core of this method lies in processing the image using a denoising model and a semantic segmentation model, calculating noise scores for normal and abnormal regions separately, and finally combining the scores of these two regions to obtain the overall noise evaluation result. After inputting the image to be tested, it is denoised using a denoising model to generate a denoised image. The denoising model used in this technique is the Restormer network. The purpose of denoising is to remove noise from the image. Then, an index is calculated between the denoised image and the original image. The underlying idea is that the smaller the difference between the two images, the lower the noise level of the original image, and the higher the score.

[0138] This technique uses a semantic segmentation model to process the test image and generate semantic segmentation results. The BiSeNetv2 model is selected for the segmentation task. The semantic segmentation model divides the image into different regions, which are then empirically categorized into normal and abnormal regions. Semantic segmentation can more accurately identify fine-grained regions in the image, avoiding treating these details as noise during noise evaluation. Based on the semantic segmentation results, normal and abnormal region masks are generated. After generating the mask images, the normal and abnormal region masks are applied to the test image and the denoised image, respectively, to extract the original and denoised images of the normal and abnormal regions. For normal and abnormal regions, PSNR and SSIM indices are calculated to measure the noise level of the image and the similarity of the overall structure of the image before and after processing. After calculation, these values ​​are finally converted into a noise score in the range of 0 to 10 using a series of mapping methods. This score can be used as the output of the abnormal region removal calculation mode.

[0139] Furthermore, texture scores are calculated for normal and abnormal regions. Texture scores are measured by the contrast attribute obtained from calculating the gray-level co-occurrence matrix (GLCM) of the image, reflecting the complexity of details in the image. The texture scores of normal and abnormal regions are used for weighted calculations to ensure that the influence of high-texture regions is appropriately reduced. A series of processing and weighted calculations are performed, taking into account both texture and the corresponding image area proportions of the two regions, to synthesize the noise scores of normal and abnormal regions, obtaining the final noise evaluation result. This score can be used as the output of the weighted calculation mode for abnormal regions. This method not only considers the overall noise level of the image but also can meticulously evaluate the noise situation in different regions of the image, especially in images containing complex details, exhibiting higher accuracy and reasonableness. By combining a denoising model and a semantic segmentation model, this technical solution achieves more accurate and detailed image noise evaluation.

[0140] Specifically, first, select the image loading path and load the image to be tested, i.e., the original image A, from the specified path. This image will be used for subsequent noise evaluation processing.

[0141] Subsequently, a denoising model is used to denoise the image to be evaluated, resulting in a denoised image. The original image A is denoised using the Restormer denoising model to obtain denoised image B. Restormer is a deep learning-based denoising network that uses neural networks to remove noise from images, improving visual quality and providing a reference benchmark for subsequent noise evaluation. In the specific model inference process, the image to be evaluated is first loaded into memory and converted to floating-point format for neural network processing. Next, the image is padded to ensure its size is an integer multiple of a specific value (e.g., a multiple of 8) to accommodate network operations. The padded operation uses reflection padded to maintain the continuity of image edges. During processing, to achieve more detailed and effective denoising, a block-based denoising approach is adopted, dividing the image into multiple small blocks. The block size is 64×72 pixels. By traversing all small blocks and performing denoising on each block separately, it is ensured that all pixels are processed uniformly. After denoising, all small blocks are combined back into the complete image, and the result is cropped to remove the padded portion. Finally, the denoised image is converted from floating-point format back to byte format and saved to the specified path.

[0142] Then, a semantic segmentation model is used to perform semantic segmentation on the image to be evaluated to obtain a mask image of the abnormal regions. The original image A is processed using the BiSeNet v2 semantic segmentation model to generate a mask image M. This model is an efficient semantic segmentation network that can segment the image into different regions based on different semantics. Regions that are frequently identified as abnormal due to noise, such as roads and grass, are classified as abnormal regions, while other non-abnormal regions are defined as normal regions. The mask image is used to identify normal and abnormal regions in the image. In the abnormal region mask image M, a value of 0 (black) represents a normal region, and a value of 1 (white) represents an abnormal region. This step ensures that normal and abnormal regions can be processed separately in subsequent calculations, thus more accurately evaluating the image noise level. Flipping the color of M yields the normal region mask image. Using M to perform a masking operation on the test image A yields an image with abnormal regions removed.

[0143] Then, the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) are calculated using the non-abnormal regions of the denoised image and the original image under test. Before calculating the metrics, all images involved in the calculation are preprocessed to ensure a uniform image size of 1280×720 pixels. This step aims to provide consistent input conditions for subsequent image processing, ensuring the repeatability of the image processing workflow and the comparability of the results. Non-abnormal regions (i.e., the parts with a value of 0 in mask M) are extracted from the denoised image B and the original image A, and the PSNR and SSIM are calculated. The PSNR and SSIM are mapped to a 10-point scale, and then averaged to a noise score of 1 for the non-abnormal regions, which can be used as the output selection.

[0144] In addition, peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) can be calculated using the abnormal regions of the denoised image and the original image under test. Similarly, all images involved in the calculation must be preprocessed before calculation to ensure a uniform image size of 1280×720 pixels. This step aims to provide consistent input conditions for subsequent image processing, ensuring the repeatability of the image processing workflow and the comparability of the results. Abnormal regions (i.e., the parts with a value of 1 in mask M) are extracted from the denoised image B and the original image A, and PSNR and SSIM are calculated. PSNR and SSIM are mapped to a 10-point scale and then averaged to obtain the noise score for the abnormal region.

[0145] Then, the contrast ratio of the gray-level feature matrix is ​​calculated for both abnormal and non-abnormal regions of the original image to represent texture information. A weighted noise score 2 is obtained using texture information, area proportion, noise score 1 for non-abnormal regions, and noise score for abnormal regions, which can be used as another output option. Specifically, the contrast ratio of the gray-level co-occurrence matrix (GLCM) for abnormal and non-abnormal regions of the original image A is calculated separately to represent texture information. The GLCM describes the texture structure of an image by statistically analyzing the spatial co-occurrence characteristics between gray levels. Then, the texture information, region area proportion, and the previously calculated noise scores for non-abnormal and abnormal regions are combined and weighted to obtain a comprehensive noise score 2.

[0146] Through the above steps, this invention achieves the calculation of a comprehensive noise score, enabling a more accurate assessment of image noise levels. This yields the noise assessment data shown in Table 1 below, where the "Weighted Calculation of Abnormal Regions" row represents non-abnormal / abnormal regions. This clearly demonstrates how existing methods misjudge abnormal regions, leading to higher noise levels and consequently lower noise scores. This application employs a calculation model that completely removes abnormal regions or performs a weighted calculation with certain weights, resulting in a more reasonable score that better meets assessment expectations.

[0147]

[0148] Table 1 Example Noise Assessment Table This application also provides an electronic device, such as... Figure 5 As shown, the electronic device 500 includes: One or more processors 510; The memory 520 stores one or more programs that, when executed by one or more processors 510, cause the one or more processors 510 to implement the image noise evaluation method described in any of the above embodiments.

[0149] Memory 520, as a non-transitory network system, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory 520 may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory 520 may optionally include remotely located memories 520 relative to processor 510, which can be connected to processor 510 via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0150] The memory 520 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 520 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 520 and is called and executed by the processor 510.

[0151] The processor 510 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0152] In some embodiments, the electronic device further includes: Input / output interfaces are used to implement information input and output; The communication interface is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). The bus transmits information between various components of the device, such as processor 510, memory 520, input / output interfaces, and communication interfaces. The processor 510, memory 520, input / output interface, and communication interface can communicate with each other within the device via a bus.

[0153] One embodiment of this application also provides a computer-readable storage medium storing computer-executable instructions for performing an image noise evaluation method as provided in any embodiment of this application.

[0154] An embodiment of this application also provides a computer program product, including a computer program or computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer program or computer instructions from the computer-readable storage medium and executes the computer program or computer instructions, causing the computer device to perform the image noise evaluation method as provided in any embodiment of this application.

[0155] The system architecture and application scenarios described in the embodiments of this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided in the embodiments of this application. Those skilled in the art will know that as the system architecture evolves and new application scenarios emerge, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0156] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0157] In hardware implementations, the division between functional modules / units mentioned in the above description does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed collaboratively by several physical components. Some or all physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software may be distributed on a computer-readable medium, which may include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, as is known to those skilled in the art, communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0158] As used in this specification, the terms "component," "module," "system," etc., are used to refer to computer-related entities, hardware, firmware, combinations of hardware and software, software, or software in execution. For example, a component can be, but is not limited to, a process running on a processor, a processor, an object, an executable file, an execution thread, a program, or a computer. As illustrated, applications running on computing devices and computing devices can both be components. One or more components may reside in a process or execution thread, and components may be located on a single computer or distributed among two or more computers. Furthermore, these components can be executed from various computer-readable media on which various data structures are stored. Components can communicate, for example, via local or remote processes based on signals having one or more data packets (e.g., data from two components interacting with another component between a local system, a distributed system, or a network, such as the Internet interacting with other systems via signals).

[0159] The above description, with reference to the accompanying drawings, illustrates some embodiments of this application, but does not limit the scope of this application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and spirit of this application shall be within the scope of this application.

Claims

1. An image noise assessment method, comprising: Semantic segmentation is performed on the image to be evaluated to obtain a first region and a second region in the image to be evaluated. The first region is a fine-grained entity region that interferes with the identification of actual noise, and the second region is the region other than the first region. Noise assessment is performed on the target region of the image to be evaluated to obtain a noise assessment score for the image to be evaluated; The target area includes at least the second area.

2. The method according to claim 1, characterized in that, The step of performing noise evaluation on the target region of the image to be evaluated to obtain a noise evaluation score for the image to be evaluated includes: Obtain the first region and the second region in the denoised image of the image to be evaluated; The target region of the image to be evaluated is compared and analyzed with the target region of the denoised image to obtain the noise evaluation score of the image to be evaluated.

3. The method according to claim 2, characterized in that, The step of comparing and analyzing the target region of the image to be evaluated with the target region of the denoised image to obtain a noise evaluation score for the image to be evaluated includes: A first noise index is calculated between the target region of the image to be evaluated and the target region of the denoised image to obtain a noise evaluation score for the image to be evaluated. The first noise index includes at least one of peak signal-to-noise ratio and structural similarity index.

4. The method according to claim 3, characterized in that, The step of calculating a first noise index between the target region of the image to be evaluated and the target region of the denoised image to obtain a noise evaluation score for the image to be evaluated includes: A first noise index is calculated between the target region of the image to be evaluated and the target region of the denoised image to obtain the first noise index result. The first noise index result is mapped to the target scaling interval through the target mapping function to obtain the scaling noise index result; Based on the scaled noise index results, the noise evaluation score of the image to be evaluated is determined.

5. The method according to claim 4, characterized in that, The target mapping function is a logistic function composed of an exponential function with the natural constant as the base.

6. The method according to claim 2, characterized in that, Before obtaining the first region and the second region in the denoised image of the image to be evaluated, the method further includes: The image to be evaluated is divided into multiple image blocks; Each of the image blocks is denoised to obtain multiple denoised image blocks. The denoised image blocks are stitched together to obtain the denoised image of the image to be evaluated.

7. The method according to claim 1, characterized in that, The noise assessment score includes at least one of a first noise assessment score and a second noise assessment score; The first noise evaluation score is the score obtained after performing noise evaluation on the second region of the image to be evaluated; The second noise evaluation score is a weighted comprehensive score obtained by performing noise evaluations on the first region and the second region of the image to be evaluated respectively.

8. The method according to claim 7, characterized in that, The noise assessment score includes the second noise assessment score; Before performing noise evaluation on the target region of the image to be evaluated to obtain the noise evaluation score of the image to be evaluated, the method further includes: Based on the difference in a second noise index between the first region and the second region of the image to be evaluated, a first evaluation weight for the first region and a second evaluation weight for the second region are determined. The second noise metric includes at least one of texture degree and area ratio.

9. The method according to claim 8, characterized in that, The texture is characterized by the contrast of the grayscale feature matrix.

10. The method according to claim 8, characterized in that, The step of determining a first evaluation weight for the first region and a second evaluation weight for the second region based on the difference in a second noise index between a first region and a second region of the image to be evaluated includes: Based on the first texture degree of the first region and the second texture degree of the second region of the image to be evaluated, the first texture weight of the first region and the second texture weight of the second region are determined. A first initial weight is determined based on the first texture weight and the first area ratio of the first region, and a second initial weight is determined based on the second texture weight and the second area ratio of the second region. The first initial weight and the second initial weight are normalized respectively to obtain the first evaluation weight of the first region and the second evaluation weight of the second region.

11. The method according to claim 10, characterized in that, The step of determining the first texture weight of the first region and the second texture weight of the second region based on the first texture degree of the first region and the second texture degree of the second region of the image to be evaluated includes: The first texture weight of the first region is determined based on the reciprocal of the first texture degree. The second texture weight of the second region is determined based on the reciprocal of the second texture degree.

12. The method according to claim 1, characterized in that, The semantic segmentation process of the image to be evaluated, to obtain a first region and a second region in the image to be evaluated, includes: The image to be evaluated is subjected to semantic segmentation to obtain the mask image corresponding to the image to be evaluated. Based on the mask image, the first region and the second region in the image to be evaluated are determined.

13. An electronic device, comprising: At least one processor; At least one memory for storing at least one program; The image noise evaluation method according to any one of claims 1 to 12 is implemented when at least one of the programs is executed by at least one of the processors.

14. A computer-readable storage medium storing computer-executable instructions for performing the image noise assessment method according to any one of claims 1 to 12.

15. A computer program product comprising a computer program or computer instructions stored in a computer-readable storage medium, wherein a processor of an electronic device reads the computer program or computer instructions from the computer-readable storage medium, and the processor executes the computer program or computer instructions to cause the electronic device to perform the image noise evaluation method according to any one of claims 1 to 12.

Images