A Perceptual Artifact Detection and Removal Method for High Dynamic Range Reconstruction

Abstract

This invention relates to the field of HDR reconstruction, specifically to the task of reconstructing HDR images from multi-exposure LDR images. A perceptual artifact detection and elimination method for high dynamic range reconstruction is proposed, comprising the following steps: Step 1, establishing an HDR-AD dataset; Step 2, constructing and training an HDR-ADet model; building an artifact detector HDR-ADet based on ViT, and training this detector using the HDR-AD dataset to locate artifact regions in the image; Step 3, using the artifact mask obtained from HDR-ADet to fine-tune the HDR reconstruction model to improve the quality of the reconstructed HDR image; and further improving the quality of the reconstructed HDR image by integrating HDR-ADet into the loss function calculation process of the HDR reconstruction framework. This invention can improve the quality of HDR imaging, achieving breakthroughs not only in quantitative indicators but also more closely approximating human visual perception.

Description

Technical Field

[0001] This invention relates to the field of HDR reconstruction, and more specifically to the task of reconstructing HDR images from multi-exposure LDR images. Background Technology

[0002] Because camera sensors have a limited light-sensing range, images typically have a low dynamic range (LDR), which loses detail in very bright or very dark areas. Conversely, high dynamic range (HDR) images capture a wider spectrum, preserving rich detail and vibrant colors. This capability makes HDR imaging invaluable in fields such as medical imaging, remote sensing, and photography, driving the growing demand for high-quality HDR images. While some hardware can directly capture HDR images, their high cost prevents widespread adoption. A more readily achievable alternative is to merge LDR images with different exposures to create an HDR image, which expands the dynamic range and recovers more detail. However, camera or object movement often causes misalignment between LDR images, resulting in ghosting artifacts in the merged HDR image. Furthermore, saturation in extremely bright or dark areas introduces distortion artifacts, affecting image quality and reducing visual appeal.

[0003] In existing methods, convolutional neural network (CNN)-based approaches focus on fusing LDR inputs after optical flow alignment; end-to-end models introduce various attention mechanisms to suppress artifacts during the fusion phase.

[0004] Recently, transformers have proven effective in this task because they are able to capture global contextual information. One specific approach to using transformers for HDR reconstruction is to first extract coarse features from LDR images with different exposures using a spatial attention mechanism, and then use multiple context-aware transformer blocks (CTBs) for HDR reconstruction.

[0005] Existing methods attempt to address artifacts in HDR imaging by improving alignment and fusion quality. Specifically, they extract information from high- and low-exposure images to fill in missing parts of the intermediate-exposure image, implicitly mitigating artifacts. However, there is currently no method that explicitly addresses this issue by detecting and locating artifact regions in the reconstructed HDR image. Furthermore, because these artifact regions are typically small, they are often ignored in quantitative metric calculations due to the average value, yet their impact on image quality is significant. Currently, there are no specific HDR image quality evaluation metrics for artifacts.

[0006] In summary, the disadvantages of existing technologies are:

[0007] 1. There is a lack of HDR reconstruction methods that can explicitly solve the artifact problem; most methods solve it implicitly by extracting information.

[0008] 2. There is a lack of evaluation indicators that can reflect the impact of artifacts on image quality and visual experience. Summary of the Invention

[0009] To address the problems of existing technologies, this invention aims to explicitly solve the artifact problem in multi-exposure HDR reconstruction tasks by detecting and locating artifact regions in reconstructed HDR images. Since artifact regions significantly impact human visual perception, the detected artifact mask can be incorporated as human feedback into the HDR reconstruction process, thereby improving the visual quality of the reconstructed image. This artifact mask can also serve as an image evaluation metric, reflecting the visual quality of the HDR image through the size of the artifact regions.

[0010] This invention treats HDR artifacts as unique and detectable entities, addressing them explicitly rather than relying solely on alignment and fusion methods. The goal is to identify artifact regions in HDR output and feed this information back to the HDR model for improvement. To this end, the invention first collects diverse LDR image sets from various scenes and reconstructs HDR output using multiple HDR models. Then, it analyzes existing artifacts and creates detailed pixel-by-pixel annotations, ultimately forming the HDR-AD (HDR Artifacts Dataset). Using HDR-AD, the invention introduces the HDR-ADet (HDR Artifacts Detector), an artifact detector based on VisionTransformers (ViTs), which takes an HDR image as input and outputs a mask highlighting artifact regions. Finally, the HDR model is improved by feeding back detected artifacts during training to enhance quality.

[0011] Technical solution

[0012] A method for detecting and eliminating perceptual artifacts in high dynamic range reconstruction includes the following steps:

[0013] Step 1: Create an HDR-AD dataset

[0014] To detect artifact regions in HDR images, a high-quality HDR-AD dataset is first constructed: LDR images acquired under different exposures, HDR images obtained through inference using existing HDR reconstruction models, and manually annotated pixels to obtain masks. The HDR images and masks are used to train the detection model.

[0015] Step 2: Construct and train the HDR-ADet model.

[0016] The artifact detector HDR-ADet was built based on ViT and trained using the HDR-AD dataset. After training on a large amount of data, HDR-ADet can accurately locate artifact regions in images.

[0017] After training, when given an HDR image, HDR-ADet will output a binary mask representing the artifact region.

[0018] Step 3 uses the artifact mask obtained from HDR-ADet to fine-tune the HDR reconstruction model, thereby improving the quality of the reconstructed HDR image.

[0019] Because artifacts are highly correlated with human visual perception, the artifact mask obtained by HDR-ADet can be considered as human feedback on the quality of HDR image reconstruction. By integrating HDR-ADet into the loss function calculation process of the HDR reconstruction framework, the quality of the model reconstructing HDR images can be improved.

[0020] The artifact mask output by HDR-ADet can also be used as an evaluation metric for HDR images: existing HDR image evaluation metrics cannot reflect the impact of artifacts on image quality. This invention proposes a novel image evaluation metric, the Artifacts Score (AS) obtained through the artifact mask, which can accurately reflect human perception of image quality.

[0021] Beneficial effects

[0022] Through the above technical solution, the present invention can successfully detect artifact regions in HDR images obtained by multiple HDR reconstruction models, and use these regions as feedback to fine-tune the HDR reconstruction model, thereby improving the HDR imaging quality. It not only achieves a breakthrough in quantitative indicators, but also is closer to human visual perception. Attached Figure Description

[0023] Figure 1A schematic diagram of the overall processing flow of the method of the present invention;

[0024] Figure 2 Detailed flowchart of the method of this invention;

[0025] Figure 3 Statistical information of the dataset obtained in the embodiments of the present invention;

[0026] Figure 4 This invention provides a detailed analysis of the qualitative and quantitative aspects of HDR reconstruction models and artifact ratios.

[0027] Figure 5 The specific network architecture diagram of the artifact detector HDR-ADet of this invention;

[0028] Figure 6 Visual comparison results of embodiments of the present invention;

[0029] Figure 7 A schematic diagram of the visualization results of the image evaluation index in an embodiment of the present invention. Detailed Implementation

[0030] The technical solutions provided in this application will be further described below with reference to specific embodiments and accompanying drawings. The advantages and features of this application will become clearer from the following description.

[0031] A method for perceptual artifact detection and elimination for high dynamic range reconstruction includes the following steps: (e.g.) Figure 1 , Figure 2 )

[0032] Step 1: Create an HDR-AD dataset

[0033] Unlike existing HDR reconstruction models that typically address artifacts indirectly through fusion and alignment, this invention employs a novel approach to directly detect and resolve artifacts. Current LDR-HDR datasets are limited in size, lack diversity, and are not specifically designed for artifacts.

[0034] To address this gap, we first construct a comprehensive HDR artifact dataset based on real-world scenes, including the following steps:

[0035] Step 1.1 Acquire a set of multi-exposure LDR images;

[0036] This invention first uses mobile phones and digital cameras to capture various image contents to support HDR across multiple platforms. All images are initially captured in RAW format at the original resolution of each device and then converted to TIFF format for further processing. To focus on challenging scenes in the HDR image generation process, this invention archives and organizes the collected LDR image sets to ensure that each final set of LDR images contains different exposures and objects with movement.

[0037] Step 1.2 Reconstructing HDR images using existing HDR reconstruction models;

[0038] Next, this invention applies mature and high-performance HDR reconstruction models (such as AHDR, HDR-Transformer, and SCTNet) to generate HDR images. The sources of the generated images include LDR images collected using the above process, as well as existing Kal's and Tel's images.

[0039] Step 1.3 Manually label the artifact mask.

[0040] Manual annotation was performed using the online platform Labelbox. During the annotation process, images with excessively high or low reconstruction quality were excluded. If multiple annotators could not find artifact regions in the same HDR image, its reconstruction quality was considered too high; if an HDR image consisted almost entirely of artifact regions and no successfully reconstructed areas could be found, its reconstruction quality was considered too low. Images with significant disagreements among annotators were also ignored to prevent model bias.

[0041] Specifically, this invention collected 1,213 LDR image sets, each containing low, medium, and high exposure images, and obtained 1,765 HDR images with pixel-by-pixel annotations through a rigorous screening process.

[0042] In addition to the LDR image set acquired in step 1.1, this invention also processes the Kal's and Tel's datasets to increase the diversity of the datasets and promote robust generalization of the detector.

[0043] Statistical information of the dataset of this invention is as follows: Figure 3 As shown, specifically:

[0044] like Figure 3 As shown in (a), the dataset covers both indoor and outdoor scenes to cover a variety of lighting conditions.

[0045] like Figure 3 (b) shows the quantitative analysis of brightness and contrast of the dataset of the present invention and the Kal's and Tel's datasets.

[0046] like Figure 3 (c) shows examples of different time periods—morning, sunrise / sunset, and night—to capture a wide range of lighting conditions.

[0047] like Figure 3As shown in (d), the LDR image set also includes a variety of motion types that produce noticeable artifacts, forming challenging cases. This invention introduces motion of people or objects in the scene to simulate misalignment during bracketing exposures (i.e., foreground motion). Furthermore, this invention intentionally moves the camera to produce camera shake (i.e., background motion), and in some image sets, these two motion types are combined (i.e., panoramic motion). This diversity of motion types is crucial for capturing a wide range of artifact variations that significantly impact HDR quality.

[0048] Step 2: Construct and train the HDR-ADet network.

[0049] The HDR-ADet network comprises a feature backbone extraction network, a feature fusion bottleneck, and a detection head. The backbone network and head architecture perform the detection task and incorporate a fusion bottleneck to better integrate features across multiple scales. Figure 5 .

[0050] Step 2.1 Constructing a Feature Backbone Extraction Network

[0051] The feature extraction backbone network employs a hybrid window attention strategy, which combines window self-attention (WSA) with multiple cross-window block (Global Self-Attention, GSA) blocks. Specifically, given an input HDR image I, the image undergoes patch embedding and Transformer block processing. The output is the final feature map F, with a patch size of 16. This process is formally expressed as follows:

[0052] F = V(I),

[0053] Among them, V(·) is a feature extraction backbone network based on window ViT.

[0054] Step 2.2 Constructing the Feature Fusion Bottleneck

[0055] Since artifacts in HDR images vary in shape and size, this invention proposes a multi-scale feature fusion network to extract features at different scales from the output of the feature extraction backbone network. Specifically, the feature map (F) from the backbone network generates multi-scale features through a series of convolutional, pooling, and deconvolutional layers. Each scale of the feature map represents information about the image at different resolutions, thus capturing a range of features from broad, coarse information to local details. Convolutional layers are responsible for extracting local features, pooling layers help reduce spatial dimensionality and enhance focus on overall image information, while deconvolutional layers recover detailed information through upsampling operations.

[0056] p i =S i(F),

[0057] Where i takes values ​​ranging from 1, 2, 3, 4, and p i The features are represented at different scales, namely 1 / 32, 1 / 16, 1 / 8, and 1 / 4 of the feature map, respectively. i (·) is a series of convolutional, pooling, and deconvolutional layers.

[0058] However, since artifacts in HDR images often exhibit complex contextual inconsistencies and sometimes involve subtle details within the image, relying solely on local features may be insufficient to capture global information. Therefore, this invention also introduces a global information extraction mechanism. This mechanism extracts global contextual information from the feature map, specifically through global pooling or a global feature extraction module. Global information p4 captures macroscopic structure and background information across the entire image, which is particularly important for artifact detection because artifacts may exhibit different morphologies due to the influence of the global scene. The extraction process is as follows:

[0059] f g =Maxpooling(p4),

[0060] Next, the present invention will extract the global information f g Upsampling to each scale feature p i The upsampling operation transforms global information into a spatial dimension that matches the feature map at each scale, allowing global information to be effectively combined with local features. In this way, global information provides supplementary contextual information for features at each scale, helping the network better understand the overall structure of the image and improving artifact detection accuracy.

[0061] f i =U(f g )+p i ,

[0062] Where U(·) represents the upsampling operation.

[0063] Through the design of the multi-scale feature fusion network described above, this invention can not only effectively handle the differences in artifacts at different scales, but also improve the network's ability to perceive artifacts by introducing global information, making artifact detection in HDR images more accurate and reliable.

[0064] Step 2.3 Construct the detection head

[0065] To strike a balance between performance and computational cost, this invention uses an MLP decoder as the detection head. First, multi-scale features are upsampled to 1 / 4 the size of the input HDR image and then stitched together. Then, two MLP layers are used to fuse the features and generate the final prediction. The MLP decoder MLP(·) can be formulated as follows:

[0066]

[0067] in It is the predicted artifact mask.

[0068] Through this design, the MLP decoder can not only efficiently fuse features from different scales and levels, but also generate accurate artifact predictions at a lower computational cost. Compared with traditional convolutional neural network decoders, the MLP decoder has the advantages of simple structure and low computational resource requirements, making it very suitable for fast processing in practical applications while maintaining high detection accuracy.

[0069] Step 2.4 Training the HDR-ADet network

[0070] The artifact detector HDR-ADet was trained in a supervised manner using the HDR-AD dataset, with the backbone network frozen and only the parameters of the feature fusion bottleneck and the detector head updated. The training epochs were set to 200.

[0071] Step 3 integrates HDR-ADet into the HDR reconstruction framework to improve the visual quality of the reconstructed HDR image.

[0072] Human visual perception is highly sensitive to visual artifacts; therefore, the artifact mask output by HDR-ADet can provide valuable guidance for visual enhancement models (i.e., HDR models), aligning them with human perceptual responses. Intuitively, integrating HDR-ADet into the HDR reconstruction framework can improve the visual quality of reconstructed HDR images, better reflecting human perception of light and color, such as... Figure 4 As shown, (a) represents HDR images reconstructed from the same LDR image by different models, and (b) represents the average artifact ratio of various models on multiple datasets.

[0073] This invention proposes a simple, effective, and unified method to optimize loss function calculation by integrating HDR-ADet into the HDR reconstruction framework: the predicted HDR image. artifact mask Used to calculate L1 loss penalty for the corresponding region during loss calculation, the corresponding loss function formula is as follows:

[0074]

[0075] Where H is the true value, L r It is the original loss of the HDR reconstruction framework. It's a hyperparameter, L h This is a feedback fine-tuning loss. By increasing the loss penalty in this way, the HDR model can pay more attention to the poor-quality areas, i.e., the artifact areas, during the fine-tuning process, thereby improving the image quality in a targeted manner.

[0076] Compared with the prior art, the present invention has the following innovations:

[0077] (1) HDR Artifact Dataset (HDR-AD): This invention creates the first dataset for HDR artifact detection, including a set of multi-exposure LDR images, HDR images with artifacts, and pixel-by-pixel artifact masks. Based on this dataset, this invention comprehensively analyzes how different scenes and models lead to artifact formation, providing a valuable benchmark for HDR model evaluation.

[0078] (2) HDR Artifact Detector (HDR-ADet): The novel detector HDR-ADet in this invention can accurately identify artifact regions in HDR images. By incorporating human perceptual feedback, HDR-ADet allows for fine-tuning of the HDR model to improve visual quality. Currently, this invention is the first to explicitly and clearly solve the HDR artifact problem through artifact detection.

[0079] (3) Robust performance and evaluation: Extensive experiments confirmed the robustness of HDR-ADet in various scenarios and HDR reconstruction models, demonstrating its effectiveness in enhancing HDR reconstruction through intra-domain and cross-domain fine-tuning. In addition, user studies verified the reliability of AS as an evaluation metric.

[0080] Example 1

[0081] An experimental platform was set up to simulate the method of the present invention, as detailed below:

[0082] Datasets and Models

[0083] First, to perform HDR artifact detection, this invention conducted extensive experiments on the created HDR-AD dataset to verify the performance of the proposed HDR-ADet across different scenes / image content and HDR reconstruction methods. Second, to evaluate the effectiveness of the proposed human feedback mechanism, this invention integrated HDR-ADet into various HDR reconstruction models (i.e., AHDR, HDR-Trans, and SCTNet) and fine-tuned it on widely used HDR reconstruction datasets (i.e., Kal's and Tel's).

[0084] Comparison of Implementation Results

[0085] Tables 1 and 2 show the improvement effect of using HDR-ADet on different datasets for HDR reconstruction models. Among the evaluation metrics used, PSNR evaluates image quality by calculating the ratio between the maximum possible signal strength and noise (error) between the original image and the reconstructed image, with a value range of [0,∞]; SSIM is a perception-based image quality evaluation metric that aims to measure the similarity of images in terms of structural information, brightness, and contrast, with a value range of [0,1]; HDR-VDP-2 is a quality evaluation metric for high dynamic range images (HDR) that aims to predict the human eye's perception of differences in image quality, with a value range of [0,100].

[0086] For each reconstruction model, Tables 1 and 2 show the results of training the HDR reconstruction models (including AHDR, HDR-Trans, and SCTNet) on the HDR reconstruction datasets (i.e., Kal's and Tel's), including: the results of the original algorithm without any operation, the results of fine-tuning using the original algorithm settings, the fine-tuning effects of HDR-ADet* trained only on a portion of the dataset of the corresponding model, and the results of HDR-ADet trained on the full dataset.

[0087] Experimental results show that, on both datasets, the three different HDR reconstruction models achieved performance improvements after fine-tuning with HDR-ADet.

[0088] Table 1. Quantitative results of HDR-ADet as a human feedback-based library-based fine-tuned HDR reconstruction model.

[0089]

[0090] Table 2. Quantitative results of HDR-ADet as a human feedback-based cross-library fine-tuning HDR reconstruction model.

[0091]

[0092]

[0093] Figure 6 The visualizations show qualitative results before and after fine-tuning using HDR-ADet, revealing that the fine-tuned models retain more detail and exhibit better visual quality.

[0094] This invention introduces a no-reference HDR image evaluation metric—Artifacts Score (AS)—based on artifact masks generated by HDR-ADet. Unlike current HDR-specific evaluation metrics, AS highlights the impact of artifacts on HDR image quality. Since the perception of image quality by artifacts is not linearly related to their size, this invention applies a non-linear mapping function to the artifact region proportion (AR), defined as follows:

[0095] AS = min(log(1+α·AR),1),

[0096] Here, α serves as a scaling factor to control the impact of artifact region proportions. In this implementation, α is set to 10, and AS is truncated to its maximum value of 1. If the score exceeds 1, the HDR image reconstruction quality is considered very poor.

[0097] This embodiment calculates the AS of all HDR images in HDR-AD-O, such as Figure 7 As shown, image quality significantly decreases with increasing AS. The results are highly consistent with human visual perception, providing a clearer understanding of image quality based on artifacts.

[0098] This invention also included user studies to examine the correlation between AS and human perception. Spearman correlation coefficients were calculated, and participants were divided into expert and non-expert groups to improve the accuracy of image quality assessment. Quantitative results in Table 3 demonstrate a high degree of consistency between AS and human visual perception.

[0099] Table 3. User survey results on HDR-ADet as an image evaluation indicator.

[0100] Tester Same scene Different scenarios total expert 0.7109 0.6093 0.6601 Non-expert 0.6701 0.4980 0.5840

Claims

1. A method for detecting and eliminating perceptual artifacts in high dynamic range reconstruction, characterized in that, Includes the following steps: Step 1: Create an HDR-AD dataset To achieve the goal of detecting artifact regions in HDR images, a high-quality HDR-AD dataset is first constructed: LDR images are acquired under different exposures, HDR images are obtained by inference using existing HDR reconstruction models, and masks are manually annotated pixel by pixel. Step 2: Construct and train the HDR-ADet model. The artifact detector HDR-ADet was built based on ViT and trained using the HDR-AD dataset to locate artifact regions in images. Step 3: Use the artifact mask obtained from HDR-ADet to fine-tune the HDR reconstruction model to improve the quality of the reconstructed HDR image. Because artifacts are highly correlated with human visual perception, the artifact mask obtained by HDR-ADet can be considered as human feedback on the quality of HDR image reconstruction. By integrating HDR-ADet into the loss function calculation process of the HDR reconstruction framework, the quality of the model reconstructing HDR images can be improved. The HDR-ADet network comprises: a feature backbone extraction network, a feature fusion bottleneck, and a detection head; the backbone network and head architecture perform the detection task and are combined with a fusion bottleneck to better integrate features across multiple scales. The specific steps are as follows: Step 2.1 Construct a feature backbone extraction network The feature extraction backbone network employs a hybrid window attention strategy, which combines window attention with multiple cross-window blocks. Specifically, given an input HDR image I, the image undergoes patch embedding and Transformer block processing; the output is the final feature map F, with a patch size of 16. This process is formally expressed as follows: Among them, V It is a feature extraction backbone network based on window ViT; Step 2.2 Constructing Feature Fusion Bottlenecks Since artifacts in HDR images vary in shape and size, a multi-scale feature fusion network is proposed to extract features of different scales from the output of the feature extraction backbone network. Specifically, the feature maps (F) from the backbone network generate multi-scale features through a series of convolutional, pooling, and deconvolutional layers. Each scale of the feature map represents information of the image at different resolutions, thus capturing various features from large-scale coarse information to local details. Convolutional layers are responsible for extracting local features, pooling layers help reduce spatial dimensions and enhance attention to the overall information of the image, while deconvolutional layers recover detailed information through upsampling operations. Where i takes values ​​ranging from 1, 2, 3, 4. These represent features at different scales, namely 1 / 32, 1 / 16, 1 / 8, and 1 / 4 of the feature map. It consists of a series of convolutional, pooling, and deconvolutional layers; A global information extraction mechanism is introduced, which extracts global contextual information from the feature map, specifically through global pooling or a global feature extraction module; global information Captures macroscopic structure and background information across the entire image range, global information. The extraction process is as follows: Next, we will extract the global information. Upsampling to each scale feature The upsampling operation transforms global information into a spatial dimension that matches the feature map at each scale, allowing global information to be effectively combined with local features. In this way, global information provides supplementary contextual information for features at each scale, helping the network better understand the overall structure of the image and improving the accuracy of artifact detection. Among them U Indicates an upsampling operation; Step 2.3 Construct the detection head First, the multi-scale features are upsampled to 1 / 4 of the size of the input HDR image and then stitched together; Then, two MLP layers are used to fuse the features and generate the final prediction; The MLP decoder MLP(·) can be formulated as follows: in It is the predicted artifact mask; Step 2.4 Training the HDR-ADet Network The artifact detector HDR-ADet was trained in a supervised manner using the HDR-AD dataset, with the backbone network frozen and only the parameters of the feature fusion bottleneck and the detection head updated.

2. The method for detecting and eliminating perceptual artifacts for high dynamic range reconstruction as described in claim 1, characterized in that, Step 1 includes the following steps: Step 1.1 Acquire a set of multi-exposure LDR images; Various image contents were captured using mobile phones and digital cameras to support HDR across multiple platforms; all images were initially captured in RAW format at the original resolution of each device and then converted to TIFF format for further processing; in order to focus on the difficult scenes in the HDR image generation process, the collected LDR image sets were archived to ensure that each set of LDR images contained different exposures and objects with movement. Step 1.2 Reconstructing HDR images using existing HDR reconstruction models; Step 1.3 Manually label the artifact mask; The online platform Labelbox was used for manual annotation. During the annotation process, images with excessively high or low reconstruction quality were excluded. If multiple annotators could not find artifact regions in the same HDR image, it was considered to have excessively high reconstruction quality. If an HDR image was almost entirely composed of artifact regions and no successfully reconstructed regions could be found, it was considered to have excessively low reconstruction quality. Images with significant disagreements among annotators were also ignored to prevent model bias.

3. The method for detecting and eliminating perceptual artifacts for high dynamic range reconstruction as described in claim 1, characterized in that, In step 3, HDR-ADet is integrated into the HDR reconstruction framework to optimize the loss function calculation: the predicted HDR image artifact mask Used to calculate L1 loss penalty for the corresponding region during loss calculation, the corresponding loss function formula is as follows: in, It is the actual value. It is the original loss of the HDR reconstruction framework. It's a hyperparameter. It is a feedback fine-tuning loss; by increasing the loss penalty in this way, the HDR model can pay more attention to the poor quality areas, that is, the artifact areas, during the fine-tuning process, thereby improving the image quality in a targeted manner.

4. The method for detecting and eliminating perceptual artifacts for high dynamic range reconstruction as described in claim 1, characterized in that, In step 3, the HDR reconstruction framework is AHDR, HDR-Trans, or SCTNet.

5. The method for detecting and eliminating perceptual artifacts for high dynamic range reconstruction as described in claim 1, characterized in that it introduces... The no-reference HDR image evaluation metric—Artifact Score (AS)—is defined as follows: Here, α is used as a scaling factor to control the proportion of the artifact region. The impact.

Images