Image data intelligent enhancement method and system based on large model

Abstract

The invention relates to an image data intelligent enhancement method and system based on a large model, and the method comprises the following steps: S1, obtaining power transmission line hardware fitting image data collected based on an unmanned plane, carrying out the image segmentation through employing a Mask R-CNN, and removing invalid images; s2, identifying and repairing abnormal images with overexposure and underexposure by analyzing brightness histogram distribution characteristics of the images; s3, adopting a bilinear interpolation algorithm to adjust all images in the brightness correction image set into a unified target size, and adopting a perceptual hash algorithm to combine with cosine similarity calculation to identify and remove repeated images to obtain a standardized de-duplicated image set; s4, carrying out ControlNet-based conditional image generation enhancement on the standardized de-duplicated image to obtain an enhanced image set; and S5, for the enhanced image set, using a pre-training detection model to generate an initial annotation, and combining manual calibration to obtain a complete annotation data set. According to the invention, the hardware fitting damage detection precision and efficiency are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of image processing, and in particular to an intelligent image data enhancement method and system based on a large model. Background Technology

[0002] As a major clean energy base, Qinghai has established a leading new energy system in China, with new energy accounting for over 60% of installed capacity and over 90% of power generation. To promote the transmission of clean energy, Qinghai is using the Qinghai-Henan DC ultra-high-voltage transmission line to send green electricity out of the province and is focusing on developing multi-energy complementary demonstration projects involving wind, solar, and energy storage. However, the harsh climate of the Qinghai Plateau region makes transmission line fittings susceptible to corrosion and mechanical damage. Traditional manual tower climbing operations not only pose safety hazards and are inefficient, but also make it difficult to detect hidden defects.

[0003] With the advancement of smart grid construction, drone-based intelligent inspection technology has become an important means to solve the problem of damage detection in power transmission line fittings. Equipped with imaging equipment and robotic arms, drones can perform non-destructive testing on the internal damage of fittings and automatically identify defects using deep learning algorithms, significantly improving inspection efficiency and accuracy. This technology not only avoids the risks of manual tower climbing but also adapts to the complex environment of high-altitude regions, providing technical support for the construction of smart grids.

[0004] However, current image annotation primarily relies on manual annotation, which is susceptible to subjective factors, leading to inconsistencies or errors that negatively impact subsequent model training. Manual annotation is also time-consuming and labor-intensive, especially when dealing with large-scale datasets, where its efficiency falls short of requirements. In image enhancement, traditional methods using transform functions cannot customize modifications to specific image attributes, and users need to manually adjust multiple parameters to achieve optimal results, resulting in limited versatility. Summary of the Invention

[0005] To address the aforementioned issues, the present invention aims to provide an intelligent image data enhancement method and system based on a large model, which effectively improves the accuracy and efficiency of hardware damage detection and meets the needs for efficient and intelligent detection in complex environments.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: The image data intelligent enhancement method based on large models includes the following steps: S1: Acquire image data of power transmission line fittings based on UAV data collection, use Mask R-CNN for image segmentation, remove invalid images, and obtain a clear and valid image set; S2: Based on a clear and effective image set, by analyzing the brightness histogram distribution characteristics of the images, abnormal images that are overexposed or underexposed are identified and repaired to obtain a brightness-corrected image set; S3: The bilinear interpolation algorithm is used to adjust all images in the brightness correction image set to a uniform target size, and the perceptual hashing algorithm combined with cosine similarity calculation is used to identify and remove duplicate images, resulting in a standardized deduplicated image set. S4: Perform conditional image generation enhancement based on ControlNet on the standardized and deduplicated images to obtain the enhanced image set; S5: For the enhanced image set, use a pre-trained detection model to generate initial annotations, and combine this with manual calibration to obtain a complete annotation dataset.

[0007] Furthermore, Mask R-CNN is used for image segmentation to remove invalid images and obtain a clear and valid image set, as follows: Mask R-CNN is based on the Faster R-CNN architecture. Through an enhanced feature pyramid network and RoIAlign alignment mechanism, it can simultaneously perform three tasks: object detection, bounding box regression, and pixel-level instance segmentation. It generates an accurate segmentation mask for each detected object and sets confidence thresholds and minimum area thresholds as filtering criteria to automatically remove detection results with too low confidence or too small area, thereby effectively removing background noise and irrelevant regions. For blurry images, the Laplacian operator is used to calculate the image sharpness. When the blurness is below a threshold... At that time, images marked as blurry are removed or processed: ; Where VarLaplace(I) is the Laplacian variance of image I; The Laplacian operator response of image I; This is the variance formula.

[0008] Furthermore, the system identifies and repairs overexposed and underexposed images, specifically as follows: Overexposed image repair employs an improved histogram equalization algorithm, which redistributes the pixel intensity values ​​of the image to achieve a more uniform brightness distribution and effectively restores the details of overexposed areas. First, the cumulative distribution function of the image is calculated, and then the original pixel values ​​are mapped to a new intensity range according to the equalization formula. Nonlinear stretching is specifically applied to the high brightness range to restore the texture and edge features obscured by overexposure. Underexposed image enhancement uses adaptive Gamma correction technology, which adjusts the Gamma parameter according to the average brightness of the image to nonlinearly brighten dark areas while maintaining the sense of layering in bright areas.

[0009] Furthermore, the standardized, deduplicated images are enhanced using ControlNet-based conditional image generation to obtain the enhanced image set, as follows: First, various conditional features are extracted from the standardized image, including Canny edge detection, depth estimation, and pose keypoint detection, to generate the corresponding control condition map. The standardized image is mapped to a low-dimensional latent space through the encoder of a variational autoencoder. In the low-dimensional latent space, a forward diffusion process is performed, gradually adding Gaussian noise to the latent features according to a preset noise schedule until it becomes completely random noise. The trained U-Net neural network predicts and removes noise from latent features. The U-Net combines text prompts and conditional control information provided by ControlNet to accurately predict the noise components to be removed in each denoising step. The backsampling process adopts the DDPM scheduling strategy to gradually restore clear latent features through multiple iterations. Finally, the latent features are restored into an enhanced image set through the VAE decoder.

[0010] Furthermore, automatic pre-labeling is used to initially label the enhanced image using a pre-trained YOLO object detection model. The YOLO object detection model identifies the target category and generates bounding box coordinates, category labels, and confidence scores. A confidence threshold is set to filter low-quality prediction results, and the non-maximum suppression (NMS) algorithm is applied to remove overlapping duplicate detection boxes, ensuring that each target has only one optimal label box.

[0011] An intelligent image data enhancement system based on a large model includes an image cleaning module, an image correction module, an image standardization module, an image enhancement module, and an image annotation module. The image cleaning module uses Mask R-CNN for image segmentation, removing invalid images to obtain a clear and valid image set. The image correction module analyzes the brightness histogram distribution characteristics of the images to identify and repair overexposed and underexposed abnormal images, obtaining a brightness-corrected image set. The image standardization module uses bilinear interpolation to adjust all images in the brightness-corrected image set to a uniform target size and employs a perceptual hashing algorithm combined with cosine similarity calculation to identify and remove duplicate images, obtaining a standardized deduplicated image set. The image enhancement module performs conditional image generation enhancement based on ControlNet on the standardized deduplicated images, obtaining an enhanced image set. The image annotation module uses a pre-trained detection model to generate initial annotations for the enhanced image set and combines this with manual calibration to obtain a complete labeled dataset.

[0012] Furthermore, Mask R-CNN is used for image segmentation to remove invalid images and obtain a clear and valid image set, as follows: Mask R-CNN is based on the Faster R-CNN architecture. Through an enhanced feature pyramid network and RoIAlign alignment mechanism, it can simultaneously perform three tasks: object detection, bounding box regression, and pixel-level instance segmentation. It generates an accurate segmentation mask for each detected object and sets confidence thresholds and minimum area thresholds as filtering criteria to automatically remove detection results with too low confidence or too small area, thereby effectively removing background noise and irrelevant regions. For blurry images, the Laplacian operator is used to calculate the image sharpness. When the blurness is below a threshold... At that time, images marked as blurry are removed or processed: ; Where VarLaplace(I) is the Laplacian variance of image I; The Laplacian operator response of image I; This is the variance formula.

[0013] Furthermore, the system identifies and repairs overexposed and underexposed images, specifically as follows: Overexposed image repair employs an improved histogram equalization algorithm, which redistributes the pixel intensity values ​​of the image to achieve a more uniform brightness distribution and effectively restores the details of overexposed areas. First, the cumulative distribution function of the image is calculated, and then the original pixel values ​​are mapped to a new intensity range according to the equalization formula. Nonlinear stretching is specifically applied to the high brightness range to restore the texture and edge features obscured by overexposure. Underexposed image enhancement uses adaptive Gamma correction technology, which adjusts the Gamma parameter according to the average brightness of the image to nonlinearly brighten dark areas while maintaining the sense of layering in bright areas.

[0014] Furthermore, the standardized, deduplicated images are enhanced using ControlNet-based conditional image generation to obtain the enhanced image set, as follows: First, various conditional features are extracted from the standardized image, including Canny edge detection, depth estimation, and pose keypoint detection, to generate the corresponding control condition map. The standardized image is mapped to a low-dimensional latent space through the encoder of a variational autoencoder. In the low-dimensional latent space, a forward diffusion process is performed, gradually adding Gaussian noise to the latent features according to a preset noise schedule until it becomes completely random noise. The trained U-Net neural network predicts and removes noise from latent features. The U-Net combines text prompts and conditional control information provided by ControlNet to accurately predict the noise components to be removed in each denoising step. The backsampling process adopts the DDPM scheduling strategy to gradually restore clear latent features through multiple iterations. Finally, the latent features are restored into an enhanced image set through the VAE decoder.

[0015] Furthermore, automatic pre-labeling is used to initially label the enhanced image using a pre-trained YOLO object detection model. The YOLO object detection model identifies the target category and generates bounding box coordinates, category labels, and confidence scores. A confidence threshold is set to filter low-quality prediction results, and the non-maximum suppression (NMS) algorithm is applied to remove overlapping duplicate detection boxes, ensuring that each target has only one optimal label box.

[0016] The present invention has the following beneficial effects: This invention addresses the problems of insufficient labeled data, poor image quality, and difficulty in defect identification in existing technologies. By combining data augmentation and deep learning methods, it improves the detectability of hardware defects in sample images, enhances the robustness of the detection algorithm, and maintains high accuracy and efficiency even in complex environments. Attached Figure Description

[0017] Figure 1 This is a structural diagram of the Mask R-CNN framework in one embodiment of the present invention; Figure 2 This is a functional structure diagram of the overall framework in one embodiment of the present invention; Figure 3 This is a schematic diagram of an image enhancement process in one embodiment of the present invention; Figure 4 This is a schematic diagram of the image annotation process in one embodiment of the present invention. Detailed Implementation

[0018] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: refer to Figure 1-4 In this embodiment, an intelligent image data enhancement method based on a large model is provided, including the following steps: S1: Acquire image data of power transmission line fittings based on UAV data collection, use Mask R-CNN for image segmentation, remove invalid images, and obtain a clear and valid image set; S2: Based on a clear and effective image set, by analyzing the brightness histogram distribution characteristics of the images, abnormal images that are overexposed or underexposed are identified and repaired to obtain a brightness-corrected image set; S3: The bilinear interpolation algorithm is used to adjust all images in the brightness correction image set to a uniform target size, and the perceptual hashing algorithm combined with cosine similarity calculation is used to identify and remove duplicate images, resulting in a standardized deduplicated image set. S4: Perform conditional image generation enhancement based on ControlNet on the standardized and deduplicated images to obtain the enhanced image set; S5: For the enhanced image set, use a pre-trained detection model to generate initial annotations, and combine this with manual calibration to obtain a complete annotation dataset.

[0019] In this embodiment, Mask R-CNN is used for image segmentation to remove invalid images and obtain a clear and valid image set, as detailed below: Mask R-CNN is based on the Faster R-CNN architecture. Through an enhanced feature pyramid network and RoIAlign alignment mechanism, it can simultaneously perform three tasks: object detection, bounding box regression, and pixel-level instance segmentation. It generates an accurate segmentation mask for each detected object and sets confidence thresholds and minimum area thresholds as filtering criteria to automatically remove detection results with too low confidence or too small area, thereby effectively removing background noise and irrelevant regions. For blurry images, the Laplacian operator is used to calculate the image sharpness. When the blurness is below a threshold... At that time, images marked as blurry are removed or processed:

[0020] Where VarLaplace(I) is the Laplacian variance of image I; The Laplacian operator response of image I; This is the variance formula.

[0021] In this embodiment, abnormal images that are overexposed or underexposed are identified and repaired, specifically as follows: Overexposed image repair employs an improved histogram equalization algorithm, which redistributes the pixel intensity values ​​of the image to make the brightness distribution more uniform and effectively restores the detail information of overexposed areas. First, the cumulative distribution function of the image is calculated, and then the original pixel values ​​are mapped to a new intensity range according to the equalization formula. Nonlinear stretching is specifically applied to the high brightness range to restore the texture and edge features obscured by overexposure. Underexposed image enhancement uses adaptive Gamma correction technology, which adjusts the Gamma parameter according to the average brightness of the image, nonlinearly brightening the dark areas while maintaining the sense of layering in the bright areas. Power function transformation effectively improves the overall contrast of the image, allowing target features hidden in shadows to be clearly presented.

[0022] In this embodiment, the standardized deduplicated images are enhanced using ControlNet-based conditional image generation to obtain an enhanced image set, as follows: First, various conditional features are extracted from the standardized image, including Canny edge detection, depth estimation, and pose keypoint detection, to generate the corresponding control condition map. The standardized image is mapped to a low-dimensional latent space through the encoder of a variational autoencoder. In the low-dimensional latent space, a forward diffusion process is performed, gradually adding Gaussian noise to the latent features according to a preset noise schedule until it becomes completely random noise. The trained U-Net neural network predicts and removes noise from latent features. The U-Net combines text prompts and conditional control information provided by ControlNet to accurately predict the noise components to be removed in each denoising step. The backsampling process adopts the DDPM scheduling strategy to gradually restore clear latent features through multiple iterations. Finally, the latent features are restored into an enhanced image set through the VAE decoder.

[0023] In this embodiment, automatic pre-labeling uses a pre-trained YOLO object detection model to perform initial labeling on the enhanced image. The YOLO object detection model identifies the target category and generates bounding box coordinates, category labels, and confidence scores. A confidence threshold is set to filter low-quality prediction results. At the same time, the non-maximum suppression (NMS) algorithm is applied to remove overlapping duplicate detection boxes, ensuring that each target has only one optimal label box.

[0024] An intelligent image data enhancement system based on a large model includes an image cleaning module, an image correction module, an image standardization module, an image enhancement module, and an image annotation module. The image cleaning module uses Mask R-CNN for image segmentation, removing invalid images to obtain a clear and valid image set. The image correction module analyzes the brightness histogram distribution characteristics of the images to identify and repair overexposed and underexposed abnormal images, obtaining a brightness-corrected image set. The image standardization module uses bilinear interpolation to adjust all images in the brightness-corrected image set to a uniform target size and employs a perceptual hashing algorithm combined with cosine similarity calculation to identify and remove duplicate images, obtaining a standardized deduplicated image set. The image enhancement module performs conditional image generation enhancement based on ControlNet on the standardized deduplicated images, obtaining an enhanced image set. The image annotation module uses a pre-trained detection model to generate initial annotations for the enhanced image set and combines this with manual calibration to obtain a complete labeled dataset.

[0025] In this embodiment, Mask R-CNN is used for image segmentation to remove invalid images and obtain a clear and valid image set, as detailed below: Mask R-CNN is based on the Faster R-CNN architecture. Through an enhanced feature pyramid network and RoIAlign alignment mechanism, it can simultaneously perform three tasks: object detection, bounding box regression, and pixel-level instance segmentation. It generates an accurate segmentation mask for each detected object and sets confidence thresholds and minimum area thresholds as filtering criteria to automatically remove detection results with too low confidence or too small area, thereby effectively removing background noise and irrelevant regions. For blurry images, the Laplacian operator is used to calculate the image sharpness. When the blurness is below a threshold... At that time, images marked as blurry are removed or processed:

[0026] Where VarLaplace(I) is the Laplacian variance of image I; The Laplacian operator response of image I; This is the variance formula.

[0027] In this embodiment, abnormal images that are overexposed or underexposed are identified and repaired, specifically as follows: Overexposed image repair employs an improved histogram equalization algorithm, which redistributes the pixel intensity values ​​of the image to make the brightness distribution more uniform and effectively restores the detail information of overexposed areas. First, the cumulative distribution function of the image is calculated, and then the original pixel values ​​are mapped to a new intensity range according to the equalization formula. Nonlinear stretching is specifically applied to the high brightness range to restore the texture and edge features obscured by overexposure. Underexposed image enhancement uses adaptive Gamma correction technology, which adjusts the Gamma parameter according to the average brightness of the image, nonlinearly brightening the dark areas while maintaining the sense of layering in the bright areas. Power function transformation effectively improves the overall contrast of the image, allowing target features hidden in shadows to be clearly presented.

[0028] In this embodiment, the standardized deduplicated images are enhanced using ControlNet-based conditional image generation to obtain an enhanced image set, as follows: First, various conditional features are extracted from the standardized image, including Canny edge detection, depth estimation, and pose keypoint detection, to generate the corresponding control condition map. The standardized image is mapped to a low-dimensional latent space through the encoder of a variational autoencoder. In the low-dimensional latent space, a forward diffusion process is performed, gradually adding Gaussian noise to the latent features according to a preset noise schedule until it becomes completely random noise. The trained U-Net neural network predicts and removes noise from latent features. The U-Net combines text prompts and conditional control information provided by ControlNet to accurately predict the noise components to be removed in each denoising step. The backsampling process adopts the DDPM scheduling strategy to gradually restore clear latent features through multiple iterations. Finally, the latent features are restored into an enhanced image set through the VAE decoder.

[0029] In this embodiment, automatic pre-annotation uses a pre-trained YOLO object detection model to perform initial annotation on the enhanced image. The YOLO object detection model identifies the target category and generates bounding box coordinates, category labels, and confidence scores. A confidence threshold is set to filter low-quality prediction results. At the same time, the Non-Maximum Suppression (NMS) algorithm is applied to remove overlapping duplicate detection boxes, ensuring that each target has only one optimal annotation box. Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0030] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0031] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0032] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0033] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An intelligent image data enhancement method based on a large model, characterized in that, Includes the following steps: S1: Acquire image data of power transmission line fittings based on UAV data collection, use Mask R-CNN for image segmentation, remove invalid images, and obtain a clear and valid image set; S2: Based on a clear and effective image set, by analyzing the brightness histogram distribution characteristics of the images, abnormal images that are overexposed or underexposed are identified and repaired to obtain a brightness-corrected image set; S3: The bilinear interpolation algorithm is used to adjust all images in the brightness correction image set to a uniform target size, and the perceptual hashing algorithm combined with cosine similarity calculation is used to identify and remove duplicate images, resulting in a standardized deduplicated image set. S4: Perform conditional image generation enhancement based on ControlNet on the standardized and deduplicated images to obtain the enhanced image set; S5: For the enhanced image set, use a pre-trained detection model to generate initial annotations, and combine this with manual calibration to obtain a complete annotation dataset.

2. The intelligent image data enhancement method based on a large model according to claim 1, characterized in that, The Mask R-CNN method is used for image segmentation to remove invalid images and obtain a clear and valid image set, as detailed below: Mask R-CNN is based on the Faster R-CNN architecture. Through an enhanced feature pyramid network and RoIAlign alignment mechanism, it can simultaneously perform three tasks: object detection, bounding box regression, and pixel-level instance segmentation. It generates an accurate segmentation mask for each detected object and sets confidence thresholds and minimum area thresholds as filtering criteria to automatically remove detection results with too low confidence or too small area, thereby effectively removing background noise and irrelevant regions. For blurry images, the Laplacian operator is used to calculate the image sharpness. When the blurness is below a threshold... At that time, images marked as blurry are removed or processed: ； Where VarLaplace(I) is the Laplacian variance of image I; The Laplacian operator response of image I; This is the variance formula.

3. The intelligent image data enhancement method based on a large model according to claim 2, characterized in that, The identification and repair of abnormal images that are overexposed and underexposed is as follows: Overexposed image repair uses an improved histogram equalization algorithm, which redistributes the pixel intensity values ​​of the image to make the brightness distribution more uniform and effectively restore the detail information of the overexposed areas; First, the cumulative distribution function of the image is calculated, and then the original pixel values ​​are mapped to a new intensity range according to the equalization formula, especially for the high brightness range, nonlinear stretching is performed to restore the texture and edge features that are covered by overexposure; Underexposed image enhancement employs adaptive Gamma correction technology, which adjusts the Gamma parameter based on the average brightness of the image to non-linearly brighten dark areas while maintaining the sense of detail in bright areas.

4. The intelligent image data enhancement method based on a large model according to claim 1, characterized in that, The standardized, deduplicated images are then subjected to ControlNet-based conditional image generation enhancement to obtain an enhanced image set, as detailed below: First, various conditional features are extracted from the standardized image, including Canny edge detection, depth estimation, and pose keypoint detection, to generate the corresponding control condition map. The standardized image is mapped to a low-dimensional latent space through the encoder of a variational autoencoder. In the low-dimensional latent space, a forward diffusion process is performed, gradually adding Gaussian noise to the latent features according to a preset noise schedule until it becomes completely random noise. The trained U-Net neural network predicts and removes noise from latent features. The U-Net combines text prompts and conditional control information provided by ControlNet to accurately predict the noise components to be removed in each denoising step. The backsampling process adopts the DDPM scheduling strategy to gradually restore clear latent features through multiple iterations. Finally, the latent features are restored into an enhanced image set through the VAE decoder.

5. The intelligent image data enhancement method based on a large model according to claim 1, characterized in that, The automatic pre-labeling uses a pre-trained YOLO object detection model to perform initial labeling on the enhanced image. The YOLO object detection model identifies the target category and generates bounding box coordinates, category labels, and confidence scores. A confidence threshold is set to filter low-quality prediction results. At the same time, the non-maximum suppression (NMS) algorithm is applied to remove overlapping duplicate detection boxes, ensuring that each target has only one optimal label box.

6. An intelligent image data enhancement system based on a large model, comprising an image cleaning module, an image correction module, an image standardization module, an image enhancement module, and an image annotation module; the image cleaning module uses Mask R-CNN for image segmentation, removes invalid images, and obtains a clear and valid image set; the image correction module analyzes the brightness histogram distribution characteristics of the images to identify and repair overexposed and underexposed abnormal images, obtaining a brightness-corrected image set; the image standardization module uses a bilinear interpolation algorithm to adjust all images in the brightness-corrected image set to a uniform target size, and uses a perceptual hashing algorithm combined with cosine similarity calculation to identify and remove duplicate images, obtaining a standardized deduplicated image set; the image enhancement module performs conditional image generation enhancement based on ControlNet on the standardized deduplicated images, obtaining an enhanced image set; the image annotation module uses a pre-trained detection model to generate initial annotations for the enhanced image set, and combines this with manual calibration to obtain a complete annotation dataset.

7. The image data intelligent enhancement system based on a large model according to claim 6, characterized in that, The Mask R-CNN method is used for image segmentation to remove invalid images and obtain a clear and valid image set, as detailed below: Mask R-CNN is based on the Faster R-CNN architecture. Through an enhanced feature pyramid network and RoIAlign alignment mechanism, it can simultaneously perform three tasks: object detection, bounding box regression, and pixel-level instance segmentation. It generates an accurate segmentation mask for each detected object and sets confidence thresholds and minimum area thresholds as filtering criteria to automatically remove detection results with too low confidence or too small area, thereby effectively removing background noise and irrelevant regions. For blurry images, the Laplacian operator is used to calculate the image sharpness. When the blurness is below a threshold... At that time, images marked as blurry are removed or processed: ； Where VarLaplace(I) is the Laplacian variance of image I; The Laplacian operator response of image I; This is the variance formula.

8. The image data intelligent enhancement system based on a large model according to claim 7, characterized in that, The identification and repair of abnormal images that are overexposed and underexposed is as follows: Overexposed image repair uses an improved histogram equalization algorithm, which redistributes the pixel intensity values ​​of the image to make the brightness distribution more uniform and effectively restore the detail information of the overexposed areas; First, the cumulative distribution function of the image is calculated, and then the original pixel values ​​are mapped to a new intensity range according to the equalization formula, especially for the high brightness range, nonlinear stretching is performed to restore the texture and edge features that are covered by overexposure; Underexposed image enhancement employs adaptive Gamma correction technology, which adjusts the Gamma parameter based on the average brightness of the image to non-linearly brighten dark areas while maintaining the sense of detail in bright areas.

9. The intelligent image data enhancement system based on a large model according to claim 6, characterized in that, The standardized, deduplicated images are then subjected to ControlNet-based conditional image generation enhancement to obtain an enhanced image set, as detailed below: First, various conditional features are extracted from the standardized image, including Canny edge detection, depth estimation, and pose keypoint detection, to generate the corresponding control condition map. The standardized image is mapped to a low-dimensional latent space through the encoder of a variational autoencoder. In the low-dimensional latent space, a forward diffusion process is performed, gradually adding Gaussian noise to the latent features according to a preset noise schedule until it becomes completely random noise. The trained U-Net neural network predicts and removes noise from latent features. The U-Net combines text prompts and conditional control information provided by ControlNet to accurately predict the noise components to be removed in each denoising step. The backsampling process adopts the DDPM scheduling strategy to gradually restore clear latent features through multiple iterations. Finally, the latent features are restored into an enhanced image set through the VAE decoder.

10. The image data intelligent enhancement system based on a large model according to claim 6, characterized in that, The automatic pre-labeling uses a pre-trained YOLO object detection model to perform initial labeling on the enhanced image. The YOLO object detection model identifies the target category and generates bounding box coordinates, category labels, and confidence scores. A confidence threshold is set to filter low-quality prediction results. At the same time, the non-maximum suppression (NMS) algorithm is applied to remove overlapping duplicate detection boxes, ensuring that each target has only one optimal label box.

Images