Image processing analysis system based on AI and visual inspection technology

Abstract

The invention relates to the technical field of image processing, in particular to an image processing analysis system based on AI and visual inspection technology, comprising: an image acquisition module for acquiring old pictures, the image acquisition module comprising a professional flat scanner and a macro camera; the preprocessing module is connected with the image acquisition module; according to the scheme, high-quality image data are obtained through high-precision image acquisition equipment and an advanced preprocessing algorithm, damage, blurring and other problems are intelligently repaired by means of a deep learning model in the AI algorithm module, and accurate flaw positioning and scientific repair effect evaluation are achieved through the visual detection module. And the optimal repair strategy is automatically selected through multi-module cooperative work. According to the system, high-quality restoration and efficient processing of old pictures are realized, manual intervention is greatly reduced, the stability and reliability of renovation quality are improved, and the actual requirements of scenes such as cultural heritage protection and family image renovation are met.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to an image processing and analysis system based on AI and visual detection technology. Background Art

[0002] The renovation of old pictures is an important means of repairing and optimizing precious images that carry historical memories and cultural values. It is of great significance in the fields of cultural heritage protection and family image preservation. As people pay more attention to history and culture, the demand for the renovation of old pictures is growing. However, existing renovation systems mostly rely on traditional image processing technologies, such as simple filtering and denoising, manual color adjustment, etc. These systems have poor restoration effects when faced with old pictures that are severely faded and complexly damaged, and it is difficult to restore the original colors and details of the pictures. In addition, the restoration process has a low degree of automation, requires a lot of manual operations, and is inefficient, which cannot meet the needs of large-scale renovation of old pictures. At the same time, there is a lack of scientific evaluation of the restoration effect, making it difficult to ensure the stability of the renovation quality.

[0003] To this end, this paper proposes an image processing and analysis system based on AI and visual inspection technologies. This system uses high-precision image acquisition equipment and advanced preprocessing algorithms to acquire high-quality image data. The deep learning model within the AI algorithm module intelligently repairs issues such as damage and blur. The visual inspection module accurately locates defects and scientifically evaluates the effectiveness of repairs. Furthermore, through the collaborative work of multiple modules, the optimal repair strategy is automatically selected. This system achieves high-quality restoration and efficient processing of old images, significantly reducing manual intervention and improving the stability and reliability of renovation quality, meeting the practical needs of scenarios such as cultural heritage preservation and home image renovation. Summary of the Invention

[0004] Technical problems to be solved: Lack of scientific evaluation of restoration effects, making it difficult to ensure the stability of renovation quality.

[0005] In response to the shortcomings of the existing technology, the present invention provides an image processing and analysis system based on AI and visual detection technology, including:

[0006] An image acquisition module, used to acquire old pictures, comprising a professional-grade flatbed scanner and a macro camera;

[0007] A preprocessing module, connected to the image acquisition module, for performing denoising, grayscale, color correction and geometric correction on the acquired old images;

[0008] An AI algorithm module, connected to the preprocessing module, is used to perform restoration, super-resolution reconstruction, color enhancement, and style transfer on the preprocessed old images;

[0009] A visual inspection module, connected to the preprocessing module and the AI algorithm module, for detecting defects in old images and evaluating the restoration effect;

[0010] A feature extraction and fusion module connects the preprocessing module, the AI algorithm module, and the visual detection module to extract and fuse traditional features and deep learning features of old images;

[0011] A decision-making and classification module, connecting the visual inspection module and the feature extraction and fusion module, is used to formulate a repair strategy based on the inspection results and feature information and classify the repair effects;

[0012] The result output module connects the AI algorithm module and the decision and classification module to output the restored old pictures and generate a renovation report.

[0013] In one possible implementation, the AI algorithm module includes:

[0014] Generative adversarial network restoration unit, which uses generative adversarial networks to repair damaged and missing areas in old images;

[0015] Deep learning interpolation algorithm unit, which uses deep learning interpolation algorithm to perform accurate interpolation, repair scratches and creases, and restore image integrity;

[0016] A super-resolution reconstruction unit, which is used to improve the resolution of old images using a super-resolution algorithm based on a convolutional neural network or an enhanced super-resolution algorithm;

[0017] The color enhancement and style transfer unit is used to enhance the color of old images through the color enhancement network and realize artistic style transfer using the style transfer algorithm.

[0018] In one possible implementation, the visual detection module includes:

[0019] A defect detection unit, which uses threshold-based or deep learning detection methods to detect stains, mildew, scratches, and other defects in old images.

[0020] The restoration effect evaluation unit is used to evaluate the effect of restored old pictures by using objective evaluation indicators such as peak signal-to-noise ratio and structural similarity index combined with subjective evaluation.

[0021] In one possible implementation, the feature extraction and fusion module includes:

[0022] Traditional feature extraction unit, used to extract local feature points, edges and shape features of old images using SIFT, SURF and HOG algorithms;

[0023] A deep learning feature extraction unit, which uses a pre-trained convolutional neural network to extract high-level semantic features from old images;

[0024] The feature fusion unit is used to fuse traditional features and deep learning features using early fusion or late fusion.

[0025] In one possible implementation, the decision and classification module includes:

[0026] A repair strategy decision unit is used to select a repair strategy from a preset decision rule library based on the defect information detected by the visual inspection module and the feature information provided by the feature extraction and fusion module;

[0027] The restoration effect classification unit is used to classify the restored old pictures into four levels: excellent, good, average, and needs improvement, according to the preset quality grade standards, combined with objective evaluation indicators and subjective evaluation results, and feed back the pictures rated as needs improvement to the AI algorithm module for secondary restoration.

[0028] Beneficial effects compared with existing technologies:

[0029] 1. In this solution, high-quality data acquisition and optimization of old pictures are achieved through the combination of high-precision image acquisition equipment and advanced pre-processing algorithms. The system uses professional-grade flatbed scanners and macro cameras, which can accurately capture the subtle flaws and textures of old pictures. Combined with pre-processing technologies such as non-local mean denoising and color correction, it can effectively remove noise, restore color, and correct geometric distortion. This enables the AI algorithm module to perform repairs based on clear and accurate image data, such as using GAN to repair damaged areas and super-resolution reconstruction to improve clarity. The final output is a refurbished picture that is close to or even exceeds the original quality, providing high-quality image resources for cultural heritage protection and family image preservation.

[0030] 2. This solution leverages the deep integration of AI algorithms and visual inspection technology to achieve intelligent and efficient renovation of old images. The various deep learning models in the AI algorithm module precisely address specific issues, such as damage repair and color enhancement. The visual inspection module not only quickly locates defects but also evaluates the effectiveness of restoration through objective indicators and subjective evaluation. The feature extraction and fusion module integrates traditional and deep learning features, and the decision-making and classification module automatically selects restoration strategies and classification results based on this. This collaborative operation reduces manual intervention and significantly shortens renovation processing time, while ensuring the stability and reliability of restoration quality, meeting the actual needs of large-scale renovation of old images. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention and implement it according to the contents of the specification, the following is a detailed description of the preferred embodiments of the present invention with reference to the accompanying drawings.

[0032] Figure 1 Schematic diagram of the system framework of the present invention. DETAILED DESCRIPTION

[0033] Preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. However, the present invention can be implemented in various forms, so the present invention is not limited to the embodiments described below.

[0034] The technical solution in the embodiments of the present application is to solve the problems of the above-mentioned background technology, and the overall idea is as follows:

[0035] Example:

[0036] like Figure 1 As shown, this embodiment introduces an image processing and analysis system based on AI and visual inspection technology. The system is dedicated to solving the problems of fading, damage, scratches, blurring, etc. in old pictures. Through the collaborative operation of multiple modules such as image acquisition, preprocessing, AI algorithm repair, and visual inspection assistance, the system can achieve intelligent repair, color restoration, and detail enhancement of old pictures, and ultimately output high-quality refurbished pictures to meet the needs of scenarios such as cultural heritage protection and family image preservation. The details are as follows:

[0037] Image acquisition module

[0038] 1.1 Hardware equipment selection

[0039] High-precision scanners: The restoration of old images requires extremely high image capture accuracy, so a professional-grade flatbed scanner is preferred. For example, the Epson Perfection V850 Pro, with an optical resolution of up to 6400 x 9600 dpi, can capture subtle textures, fading, and damaged details in old photos. For large-format old maps, posters, etc., the Heidelberg PrimeScan series scanners are recommended, supporting A0 format scanning to ensure complete image capture. To minimize physical damage to old images during the scanning process, the scanner should be equipped with anti-static glass and a soft pressing device to prevent sticking and wrinkling of the images.

[0040] Macro camera system: For old images with minor damage, mildew, and other local defects, a macro camera is required for local high-definition capture. For example, the Canon EOS R5 paired with the MP-E 65mm f / 2.81-5x macro lens can achieve 1x to 5x magnification, clearly revealing millimeter-level or even micron-level defects in old images. At the same time, to ensure shooting stability, a professional tripod, macro head, and ring light are required to avoid image blur or shadows caused by shaking and uneven lighting.

[0041] 1.2 Installation and debugging

[0042] Scanning environment setup: The scanner should be placed in an environment with relatively stable temperature and humidity, ideally between 20°C and 25°C, and a relative humidity of 40% to 60%. This prevents temperature and humidity fluctuations from affecting old image materials and scan quality. Before scanning, gently wipe the scanner glass surface with a professional dust-free cloth to ensure there is no dust or stains that could interfere with image acquisition.

[0043] Adjust shooting parameters: When using a macro camera, adjust the shooting parameters based on the degree of damage and detail required for the old image. For minor defects, set the aperture to f / 8-f / 11 to ensure sufficient depth of field. Adjust the shutter speed based on the fill light to avoid blur caused by hand shaking or subject movement. At the same time, use the live view function to zoom in on a part of the image and fine-tune the camera position and focus to ensure that the defective area is clearly in focus.

[0044] Data transmission settings: Image data collected by scanners and cameras is transferred to the computer via a high-speed USB 3.0 interface or an SD card. For large-capacity image files, wired network transmission or NAS (Network Attached Storage) devices can be used to ensure data transmission stability and efficiency and avoid data loss due to transmission interruptions.

[0045] 2. Preprocessing module

[0046] 2.1 Denoising Processing Unit

[0047] Non-local means (NLM) denoising: Old images are often noisy due to their age. The NLM algorithm effectively removes noise by searching for similar local regions within the image and taking a weighted average of the pixel values in these regions. During implementation, appropriate search window sizes (e.g., 11×11) and matching window sizes (e.g., 5×5) are set. A large window size affects processing speed, while a small window size may prevent accurate detection of similar regions. To address the granular noise common in old photos, the algorithm adjusts the smoothing parameters to remove noise while preserving image detail to the greatest extent possible.

[0048] Bilateral filtering: Bilateral filtering combines the spatial proximity and pixel value similarity of an image to effectively preserve edges while removing noise. It can also avoid blurring the boundaries between faded and normal areas in older images. The standard deviation of the spatial and range Gaussian kernels is adjusted based on the noise intensity and detail complexity of the image to balance denoising and detail preservation.

[0049] 2.2 Grayscale and color correction unit

[0050] Adaptive grayscale conversion: For older images with severe fading and color loss, we use an adaptive grayscale conversion method. By analyzing the brightness and contrast of different image regions, we dynamically adjust grayscale conversion parameters to better preserve detail and layering in the grayscaled image. For example, for older portraits, we enhance the contrast of the facial area to highlight facial features.

[0051] Color Correction: Using a color transfer algorithm, we extract color information from a reference image with normal color and transfer it to the old image. First, we perform a Lab color space conversion on the old image and the reference image to separate the luminance channel and the color channel. Then, we apply the color channel statistics (such as mean and standard deviation) of the reference image to the color channel of the old image to achieve color restoration. For old photos with yellowing, we can adjust the intensity of the yellow channel to remove the yellow tint and restore the true color.

[0052] 2.3 Geometric Correction Unit

[0053] Perspective Correction: Old images may become warped or deformed during storage, resulting in perspective distortion. Hough transform is used to detect straight line features in the image, such as photo borders and building outlines, and determine the image's perspective transformation parameters. Using the perspective transformation matrix, the deformed image is corrected to a normal perspective, restoring the image's geometric shape.

[0054] Registration and Correction: For old images that require stitching together, or for comparing and renovating old images of the same scene from different periods, we use feature point matching algorithms (such as SIFT and SURF) for image registration. We extract feature points from the images, calculate the transformation matrix based on the matching relationship between the feature points, and align the different images to ensure the accuracy of the stitching or comparison.

[0055] 3. AI algorithm module

[0056] 3.1. Generative Adversarial Network Restoration Unit, which uses a generative adversarial network to repair damaged and missing areas in old images;

[0057] For damaged and missing areas in old images, a GAN-based image restoration algorithm is used. A generator and discriminator network is constructed. The generator is responsible for generating restoration content based on the image information around the damaged area, while the discriminator determines whether the generated content is authentic. During the training process, a large number of old images containing damaged areas and corresponding restored images are used as training data. The generator continuously learns how to generate reasonable restoration content, and the discriminator improves its ability to distinguish between real and fake images. For example, for old photos with holes, the generator can produce a natural restoration effect based on the surrounding texture and color information.

[0058] 3.2. Deep learning interpolation algorithm unit: uses deep learning interpolation algorithm to perform accurate interpolation, repair scratches and creases, and restore image integrity;

[0059] For missing pixels in old images caused by scratches and creases, a deep learning interpolation algorithm is used. A convolutional neural network is trained to use the pixels around the missing area as input to predict the value of the missing pixels. By learning the pixel distribution patterns in different scenarios through a large amount of training data, the algorithm can accurately interpolate, repair scratches and creases, and restore the integrity of the image.

[0060] Deep learning interpolation algorithms are usually implemented based on convolutional neural networks (CNNs). Taking the CNN-based single image interpolation algorithm as an example, its network structure generally contains multiple convolution layers and activation functions. Assuming that the input image is I, after a series of convolution operations C1, C2, ..., C n And activation function σ (such as ReLU function σ(x) = max(0,x), predict the missing pixel value to get the output image I'; its simplified mathematical expression can be: I' = σ(C n (…(σ(C2(σ(C1(I))))))); In the actual training process, by minimizing the loss function (such as the mean square error loss function Where N is the number of training samples, y i is the true value, is the predicted value) to optimize the network parameters;

[0061] Example: For an old photo with missing pixels due to scratches, the scratches and surrounding pixels are used as input and processed by a trained deep learning interpolation network. The network learns the distribution of pixels in normal areas and predicts and fills in the missing pixels at the scratches. For example, in an old photo of a person's face with a scratch, the network predicts and fills in the pixels at the scratches based on pixel information such as the texture and color of the surrounding skin, restoring the face to a complete and more natural appearance.

[0062] 3.3. Super-resolution reconstruction unit, used to improve the resolution of old images using a super-resolution algorithm based on a convolutional neural network or an enhanced super-resolution algorithm;

[0063] Super-resolution convolutional neural network (SRCNN): Old images often suffer from low resolution and blurred details. SRCNN achieves super-resolution image reconstruction through three convolutional layers. The first convolutional layer extracts the underlying features of the image, the second convolutional layer performs nonlinear mapping on these features, and the third convolutional layer maps the features back to the high-resolution image. During training, high-resolution images and their corresponding low-resolution versions are used as training data. By minimizing the mean squared error between the reconstructed image and the original high-resolution image, the network parameters are optimized to improve the image resolution and clarity.

[0064] Enhanced Super-Resolution Algorithm (ESRGAN): ESRGAN builds on SRCNN by introducing a generative adversarial network and residual network architecture. The generator extracts richer image features through the residual network to generate high-resolution images. The discriminator determines the authenticity of the generated images. Furthermore, the discriminator uses a perceptual loss function that considers not only pixel differences but also semantic and perceptual information, making the reconstructed images closer to real high-resolution images in terms of detail and visual quality, effectively improving the clarity and sharpness of old images.

[0065] ESRGAN is based on the Generative Adversarial Network (GAN) and Residual Network (ResNet) structure; the generator extracts image features through the residual network. Assume that the generator is G and the input is a low-resolution image I. LR , generate high-resolution image I SR =G(I LR ). The discriminator D is used to determine whether the generated image is real. Its output is a probability value, which indicates the possibility that the image is real. At the same time, ESRGAN adopts the perceptual loss function L perceptual , which combines content loss and adversarial loss; content loss is generally calculated based on the difference of feature maps, such as the mean square error loss based on the VGG network feature map, assuming that the features extracted from the VGG network are Content loss Where M is the number of feature map elements, I HR is a real high-resolution image; the adversarial loss is L adversarial =-E x～Pdata [log(D(x))]-E z～Pz [log(1-D(G(z)))], perceptual loss L perceptual =L perceptual +γL adversarial , γ is the balance coefficient.

[0066] For example, a low-resolution, old landscape photo with blurred details is fed into the ESRGAN model. The generator uses a residual network to extract rich image features and generate a high-resolution image, making previously blurry details such as mountains and trees clearer and richer in texture. The discriminator determines the authenticity of the generated image, prompting the generator to continuously optimize. The resulting output image is closer to a realistic high-resolution landscape photo in terms of detail and visual quality. For example, previously blurry leaves become clearly discernible, and the clouds in the sky appear more detailed.

[0067] 3.4. Color enhancement and style transfer unit, used to enhance the color of old images through the color enhancement network and achieve artistic style transfer using the style transfer algorithm;

[0068] Color Enhancement Network: A specialized convolutional neural network is trained for color enhancement in old images. The network takes old images as input and outputs color-enhanced images. During training, it combines color statistics with visual perception metrics such as color contrast and saturation balance to guide the network in learning how to enhance color vividness and realism. For example, for old, dull landscape photos, the color enhancement network can make the sky bluer and the vegetation greener, restoring natural colors.

[0069] Style transfer: For renovating old images with specific artistic styles, we use a style transfer algorithm. We transfer the artistic styles of famous painters (such as Van Gogh's oil painting style and Monet's Impressionist style) to old images. By calculating the difference between the feature maps of the style image and the content image, we adjust the features of the old image so that it presents the target artistic style while retaining the original content, giving the old image new artistic value.

[0070] 4. Visual inspection module

[0071] 4.1 Defect Detection Unit

[0072] Threshold-based defect detection: For defects such as stains and mildew in old images, threshold segmentation is performed using the image's grayscale or color value differences. First, the image's grayscale histogram or color histogram is calculated to determine an appropriate threshold. Then, areas in the image with pixel values above or below the threshold are marked as defective. For grayscale images, if the grayscale value of the stained area is significantly lower than that of the surrounding normal area, the stain can be detected by setting a lower grayscale threshold.

[0073] Deep learning defect detection: A convolutional neural network is trained to detect defects in old images. Using a large number of old images with labeled defect locations and types as training data, the network learns the characteristic patterns of defects. During detection, the image to be tested is fed into the network, which then outputs the location and type of defects. This allows the network to accurately detect defects such as tiny scratches and damage that are difficult to detect using traditional methods.

[0074] 4.2. Restoration Effect Evaluation Unit

[0075] Objective evaluation metrics: Peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) are used to evaluate the restoration effect of old images. PSNR measures the degree of image distortion by calculating the pixel error between the restored image and the original high-quality image. SSIM assesses image similarity in terms of structure, brightness, and contrast. By calculating these metrics, the image quality before and after restoration is compared to quantitatively evaluate the restoration effect.

[0076] Subjective evaluation: Organize professionals and ordinary users to conduct subjective evaluations of restored old images. Develop evaluation criteria, such as color reproduction, detail clarity, and the naturalness of defect restoration, and have evaluators score and provide feedback on the restoration results based on these criteria. Combining subjective and objective evaluation indicators, we comprehensively assess the quality of the restoration of old images, providing a basis for further optimizing the restoration algorithm.

[0077] 5. Feature extraction and fusion module

[0078] 5.1 Feature Extraction Unit

[0079] Traditional feature extraction: Use algorithms such as SIFT and SURF to extract local feature points from old images. These feature points contain information about the image's scale, rotation, and illumination invariance, and can be used for image registration and defect location. At the same time, the HOG (Histogram of Oriented Gradients) algorithm is used to extract edge and shape features of the image, providing a basis for subsequent defect detection and restoration effect evaluation.

[0080] 5.2 Deep Learning Feature Extraction Unit

[0081] Use pre-trained convolutional neural networks (such as VGG and ResNet) to extract high-level semantic features from old images. These features contain semantic information and abstract features of the image, which are important for understanding image content, color enhancement, and style transfer. By inputting the image into the network, feature maps at different levels are obtained, and deep features of the image are extracted.

[0082] Take the pre-trained VGG16 network to extract image features as an example. The VGG16 network contains multiple convolutional layers and pooling layers. Assume that the input image is I, after a series of convolution operations C1, C2, ..., C n and pooling operations P1, P2, ..., P m , and obtain feature maps of different levels; its mathematical expression can be simply expressed as: F = P m (C n(…(P1(C1(I))))), where f is the extracted feature map; in practical applications, the output of a specific layer in the network is usually selected as the feature representation of the image;

[0083] Example: When performing color enhancement and style transfer on old images, the old images are input into a pre-trained VGG16 network. Feature maps such as the conv3_3 layer in the network are obtained, which contain semantic information and abstract features of the image. For example, when performing style transfer, the difference between the style image and the feature map of a specific layer of the VGG16 network is calculated, and then the features of the old image are adjusted based on these differences to make it present the target artistic style. To transfer an old photo of a person into the style of a Van Gogh painting, the VGG16 network is used to extract features from the two images. By adjusting the features of the old photo to make it closer to the features of Van Gogh's paintings, the old photo can have the stylistic characteristics of Van Gogh's paintings, such as more intense colors and more obvious brushstrokes.

[0084] 5.3 Feature Fusion Unit

[0085] Early fusion: After the image preprocessing stage, the features obtained by traditional feature extraction methods are fused with the features extracted by deep learning. For example, SIFT feature points and feature maps extracted by the VGG network are spliced together to form a new feature vector, which serves as the input of the subsequent AI algorithm module, enabling the algorithm to simultaneously utilize local detail features and high-level semantic features for repair and processing.

[0086] Late fusion: After the AI algorithm module performs image restoration and enhancement, it extracts features from the outputs of different algorithms and then fuses these features. For example, it fuses the features of the image restored by GAN with the features of the image reconstructed by super-resolution. The restoration effect is then evaluated through the visual inspection module. This combines the advantages of multiple algorithms to improve the quality of old image renovation.

[0087] 6. Decision and classification module

[0088] 6.1. Repair Strategy Decision Unit

[0089] Based on the type, location, and severity of defects detected by the visual inspection module, as well as the image feature information provided by the feature extraction and fusion module, a restoration strategy is developed. For example, for old images with large areas of damage, the GAN restoration algorithm is preferred; for low-resolution images, the super-resolution reconstruction algorithm is selected; for color issues, color enhancement or style transfer algorithms are used. By establishing a decision rule library, the restoration strategy can be automatically selected.

[0090] 6.2. Restoration Effect Classification Unit

[0091] The restored old images are classified according to their quality level into four levels: excellent, good, fair, and needs improvement. Classification criteria are formulated by combining objective evaluation indicators and subjective evaluation results. For example, images with a PSNR value higher than 40dB and an SSIM value greater than 0.9 are rated as excellent; images with a PSNR value between 35-40dB and an SSIM value between 0.8-0.9 are rated as good. For images rated as needs improvement, feedback is automatically sent to the AI algorithm module to adjust algorithm parameters or reselect the restoration algorithm for secondary restoration.

[0092] 7. Result output module

[0093] 7.1 Image Output Unit

[0094] Output restored old images in common image formats, such as JPEG, PNG, and TIFF. For images that require high-quality preservation, choose TIFF to preserve complete image information. For online sharing and general use, choose JPEG or PNG to reduce file size while maintaining a certain level of image quality. Also, provide parameter settings such as image resolution and color mode to meet the needs of different users.

[0095] 7.2 Report Generation Unit

[0096] Generates a renovation report for old images, including image comparisons before and after restoration, objective evaluation index data, restoration strategy descriptions, subjective evaluation feedback, etc.; the report is output in PDF format for user viewing and archiving, providing detailed records and references for cultural heritage protection, family image management, etc.

[0097] Finally, it should be noted that the above embodiments are merely examples for the purpose of illustrating the present invention and are not intended to limit the embodiments. Those skilled in the art will readily appreciate that other variations or modifications based on the above description are possible. It is not necessary and impossible to provide an exhaustive list of all embodiments. However, obvious variations or modifications arising therefrom remain within the scope of protection of the present invention.

Claims

1. An image processing and analysis system based on AI and visual detection technology, characterized in that: include: An image acquisition module, used to acquire old pictures, comprising a professional-grade flatbed scanner and a macro camera; A preprocessing module, connected to the image acquisition module, for performing denoising, grayscale, color correction and geometric correction on the acquired old images; An AI algorithm module, connected to the preprocessing module, is used to perform restoration, super-resolution reconstruction, color enhancement, and style transfer on the preprocessed old images; A visual inspection module, connected to the preprocessing module and the AI algorithm module, for detecting defects in old images and evaluating the restoration effect; A feature extraction and fusion module connects the preprocessing module, the AI algorithm module, and the visual detection module to extract and fuse traditional features and deep learning features of old images; A decision-making and classification module, connecting the visual inspection module and the feature extraction and fusion module, is used to formulate a repair strategy based on the inspection results and feature information and classify the repair effects; The result output module connects the AI algorithm module and the decision and classification module to output the restored old pictures and generate a renovation report.

2. The image processing and analysis system based on AI and visual detection technology according to claim 1, characterized in that: The AI algorithm module includes: Generative adversarial network restoration unit, which uses generative adversarial networks to repair damaged and missing areas in old images; Deep learning interpolation algorithm unit, which uses deep learning interpolation algorithm to perform accurate interpolation, repair scratches and creases, and restore image integrity; A super-resolution reconstruction unit, which is used to improve the resolution of old images using a super-resolution algorithm based on a convolutional neural network or an enhanced super-resolution algorithm; The color enhancement and style transfer unit is used to enhance the color of old images through the color enhancement network and realize artistic style transfer using the style transfer algorithm.

3. The image processing and analysis system based on AI and visual detection technology according to claim 1, characterized in that: The visual inspection module includes: A defect detection unit, which uses threshold-based or deep learning detection methods to detect stains, mildew, scratches, and other defects in old images. The restoration effect evaluation unit is used to evaluate the effect of restored old pictures by using objective evaluation indicators such as peak signal-to-noise ratio and structural similarity index combined with subjective evaluation.

4. The image processing and analysis system based on AI and visual detection technology according to claim 1, characterized in that: The feature extraction and fusion module includes: Traditional feature extraction unit, used to extract local feature points, edges and shape features of old images using SIFT, SURF and HOG algorithms; A deep learning feature extraction unit, which uses a pre-trained convolutional neural network to extract high-level semantic features from old images; The feature fusion unit is used to fuse traditional features and deep learning features using early fusion or late fusion.

5. The image processing and analysis system based on AI and visual detection technology according to claim 1, characterized in that: The decision and classification module includes: A repair strategy decision unit is used to select a repair strategy from a preset decision rule library based on the defect information detected by the visual inspection module and the feature information provided by the feature extraction and fusion module; The restoration effect classification unit is used to classify the restored old pictures into four levels: excellent, good, average, and needs improvement, according to the preset quality grade standards, combined with objective evaluation indicators and subjective evaluation results, and feed back the pictures rated as needs improvement to the AI algorithm module for secondary restoration.

Images