Scanning image processing method and system based on artificial intelligence

Abstract

The invention discloses a scanned image processing method and system based on artificial intelligence, and the method comprises the following steps: S1, image collection and preprocessing, S2, feature extraction and analysis, and S3, image quality restoration and optimization. The scanning image processing system based on artificial intelligence comprises an image acquisition module, a preprocessing module, a feature extraction and analysis module, a repair optimization module and a post-processing and output module. When noise is removed through the generative adversarial network, the adversarial training mechanism between the generator and the discriminator enables the generated denoised image to be close to a real noise-free image to the greatest extent, and detail information of the image is reserved; a super-resolution reconstruction algorithm combining a mean square error and perceptual loss is adopted in definition improvement, so that the image is more clear and natural visually, and accurate correction better fitting the actual situation is realized.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an artificial intelligence-based scanning image processing method and system thereof. Background Art

[0002] In today's digital age, scanned images are widely used in many fields, such as digital archiving of documents, electronic display of artworks, medical imaging records, etc.

[0003] Although the current traditional scanning image processing methods can improve some problems to a certain extent, the processing effect is often not ideal for complex and changeable scanning image situations, and the flexibility is poor, resulting in many limitations. For example, during the scanning process, uneven lighting can easily lead to problems such as shadows and poor contrast in the image; wrinkled and damaged paper documents will produce image defects after scanning, affecting subsequent recognition and use; and when faced with a large number of scanned images of different types and qualities, manual repair and optimization operations are extremely inefficient. Accordingly, the present invention proposes a scanning image processing method and system based on artificial intelligence. Summary of the invention

[0004] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a scanning image processing method and system based on artificial intelligence.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A scanning image processing method based on artificial intelligence comprises the following steps:

[0007] S1. Image acquisition and preprocessing: First, obtain the scanned image through a conventional scanning device, convert it into a common digital image format, and perform basic preprocessing operations. This includes but is not limited to grayscale processing of the image and converting the RGB value of the color image into grayscale value;

[0008] S2. Feature extraction and analysis: A deep convolutional neural network is used to extract features of the scanned image. The network structure includes multiple convolutional layers, pooling layers, and fully connected layers. The convolutional layer performs convolution operations by sliding the convolution kernel on the image to extract local features of different scales. After multiple layers of convolution and pooling operations, the extracted features are integrated and mapped to a specific feature space through the fully connected layer, which is used for subsequent classification or regression tasks to determine the type and degree of quality problems in the image.

[0009] S3. Image quality repair and optimization:

[0010] Noise removal: If noise is detected in the image, a generative adversarial network is used to generate a denoised image.

[0011] Clarity improvement: For situations where clarity is insufficient, a super-resolution reconstruction algorithm based on deep learning is used to improve image clarity by learning the mapping relationship between low-resolution images and corresponding high-resolution images;

[0012] Color correction: Use deep neural networks to learn the color transformation rules under different lighting conditions, scanning equipment and other factors, and establish the mapping relationship between the input scanned image color and the standard color space. Use a color correction model based on a convolutional neural network, and its training process optimizes network parameters by minimizing color difference loss;

[0013] Post-processing and output: Post-process the quality-restored and optimized images, such as sharpening edges to further enhance the image visual effect, and then output the processed images in the specified format for subsequent storage, viewing or other applications.

[0014] Preferably, in step S2, the convolution kernel size of the convolution layer of the deep convolutional neural network is set to 3×3 or 5×5, the step size is 1 or 2, and the filling method is SAME or VALID, and different setting combinations are used to adapt to the feature extraction requirements of different types of scanned images.

[0015] Preferably, in step S3, the super-resolution network adopted by the super-resolution reconstruction algorithm includes a plurality of residual blocks, and the network's learning and reconstruction capabilities of image features are improved through residual learning, so as to better improve the clarity of the scanned image.

[0016] Preferably, in step S3, the training sample acquisition method of the color correction model based on the convolutional neural network is to collect standard color images of the target objects in different ages and under different preservation conditions and the corresponding scanned images with color deviations, and optimize the network parameters by learning the color transformation rules between the two to achieve accurate correction of the color of the scanned images.

[0017] A scanning image processing system based on artificial intelligence, comprising:

[0018] Image acquisition module: used to connect to the scanning device, obtain the scanned original image, and transmit it to the subsequent module for processing;

[0019] Preprocessing module: receiving the image from the image acquisition module, performing preprocessing operations on the image according to the grayscale, normalization and other methods described in claim 1, and outputting the preprocessed image to the feature extraction module;

[0020] Feature extraction and analysis module: Built-in deep convolutional neural network and other related artificial intelligence algorithms to extract and analyze the features of the input pre-processed image, determine the type and degree of quality problems in the image, and pass the analysis results to the repair optimization module;

[0021] Restoration and optimization module: based on the received feature analysis results, call the corresponding artificial intelligence algorithm modules such as the denoising GAN module, super-resolution reconstruction module, color correction module, etc. as described in claim 1 to perform targeted quality restoration and optimization processing on the image, and then output the processed image to the post-processing module;

[0022] Post-processing and output module: perform post-processing operations such as sharpening on the input restored and optimized image, and output the final processed scanned image as required

[0023] Preferably, the number of convolutional layers, pooling layers and fully connected layers contained in the deep convolutional neural network in the feature extraction and analysis module and the parameters of each layer can be adjusted according to the characteristics of the scanned image in different application scenarios to achieve more accurate feature extraction and quality problem judgment.

[0024] Preferably, each artificial intelligence algorithm module in the repair optimization module can be updated by updating training data, optimizing network structure or adjusting relevant algorithm parameters to adapt to the ever-changing scanning image quality problems and the performance improvement requirements brought about by the development of artificial intelligence technology, thereby ensuring the system's continuous optimization of the scanning image quality processing effect.

[0025] The present invention has the following beneficial effects:

[0026] 1. Artificial intelligence technologies such as deep convolutional neural networks have powerful automatic feature learning capabilities. It can autonomously learn the characteristic patterns corresponding to various quality problems from a large number of scanned image samples of different quality conditions. Whether it is the subtle noise texture caused by the hardware differences of the scanning equipment, or the local clarity changes caused by the surface material and wrinkles of the scanned object, or the color deviation caused by complex ambient light, etc., the network can accurately capture and extract features. Through multi-layer convolution and pooling operations, the network can gradually abstract the high-level semantic features of the image, so as to more accurately judge the type and specific degree of quality problems in the image, and provide a reliable basis for subsequent targeted repairs.

[0027] 2. Through the collaborative work of multiple advanced artificial intelligence algorithms, different targeted algorithms are used for different types of quality problems and organically combined. When using a generative adversarial network to remove noise, the adversarial training mechanism between the generator and the discriminator makes the generated denoised image as close to the real noise-free image as possible, retaining the image's detailed information; in terms of clarity improvement, a super-resolution reconstruction algorithm combining mean square error and perceptual loss is used, which takes into account both pixel-level accuracy and the perceptual quality of the image, making the image visually clearer and more natural; for color correction, a deep neural network is used to learn complex color transformation rules to achieve more accurate correction that is more in line with the actual situation.

[0028] 3. Through artificial intelligence algorithms, it is possible to optimize the image features in depth. By learning the mapping relationship between a large number of low-resolution and high-resolution image pairs through super-resolution reconstruction algorithms, the network can understand the feature representations of different objects, textures and other elements in the image. In the process of improving the resolution, it can reasonably generate missing detail information based on these feature knowledge, making the image clarity more natural and realistic. The color correction algorithm is also based on the learning of deep features of the color space. The corrected color is more in line with the human visual perception and the true color of the actual object. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a partial code display diagram of a grayscale processing function in a scanning image processing method based on artificial intelligence proposed by the present invention;

[0030] Figure 2 This is a partial code display of a normalization processing function in a scanning image processing method based on artificial intelligence proposed by the present invention;

[0031] Figure 3 This is a partial code display of feature extraction in a scanning image processing method based on artificial intelligence proposed by the present invention;

[0032] Figure 4 This is a partial code display of the generator definition in the artificial intelligence-based scanning image processing method proposed by the present invention;

[0033] Figure 5 This is a partial code display of the discriminator definition in the artificial intelligence-based scanning image processing method proposed by the present invention;

[0034] Figure 6 This is a partial code display of a simple example loop of GAN training in an artificial intelligence-based scan image processing method proposed in the present invention. DETAILED DESCRIPTION

[0035] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0036] A scanning image processing method based on artificial intelligence comprises the following steps:

[0037] S1. Image acquisition and preprocessing: First, obtain the scanned image through a conventional scanning device, convert it into a common digital image format, and perform basic preprocessing operations. This includes but is not limited to grayscale processing of the image and converting the RGB value of the color image into grayscale value;

[0038] S2. Feature extraction and analysis: A deep convolutional neural network is used to extract features of the scanned image. The network structure includes multiple convolutional layers, pooling layers, and fully connected layers. The convolutional layer performs convolution operations by sliding the convolution kernel on the image to extract local features of different scales. After multiple layers of convolution and pooling operations, the extracted features are integrated and mapped to a specific feature space through the fully connected layer, which is used for subsequent classification or regression tasks to determine the type and degree of quality problems in the image.

[0039] S3. Image quality repair and optimization:

[0040] Noise removal: If noise is detected in the image, a generative adversarial network is used to generate a denoised image.

[0041] Clarity improvement: For situations where clarity is insufficient, a super-resolution reconstruction algorithm based on deep learning is used to improve image clarity by learning the mapping relationship between low-resolution images and corresponding high-resolution images;

[0042] Color correction: Use deep neural networks to learn the color transformation rules under different lighting conditions, scanning equipment and other factors, and establish the mapping relationship between the input scanned image color and the standard color space. Use a color correction model based on a convolutional neural network, and its training process optimizes network parameters by minimizing color difference loss;

[0043] Post-processing and output: Post-process the quality-restored and optimized images, such as sharpening edges to further enhance the image visual effect, and then output the processed images in the specified format for subsequent storage, viewing or other applications.

[0044] The convolution kernel size of the convolution layer of the deep convolutional neural network is set to 3×3 or 5×5, the step size is 1 or 2, and the padding method is SAME or VALID. Different setting combinations are used to adapt to the feature extraction requirements of different types of scanned images.

[0045] The super-resolution network used by the super-resolution reconstruction algorithm contains multiple residual blocks, which improves the network's learning and reconstruction capabilities of image features through residual learning, so as to better improve the clarity of scanned images.

[0046] The training sample acquisition method of the color correction model based on convolutional neural network is to collect standard color images of target objects from different ages and under different preservation conditions and the corresponding scanned images with color deviations. By learning the color transformation rules between the two, the network parameters are optimized to achieve accurate correction of the color of the scanned images.

[0047] A scanning image processing system based on artificial intelligence, comprising:

[0048] Image acquisition module: used to connect to the scanning device, obtain the scanned original image, and transmit it to the subsequent module for processing;

[0049] Preprocessing module: receiving the image from the image acquisition module, performing preprocessing operations on the image according to the grayscale, normalization and other methods in claim 1, and outputting the preprocessed image to the feature extraction module;

[0050] Feature extraction and analysis module: Built-in deep convolutional neural network and other related artificial intelligence algorithms to extract and analyze the features of the input pre-processed image, determine the type and degree of quality problems in the image, and pass the analysis results to the repair optimization module;

[0051] Restoration and optimization module: based on the received feature analysis results, call the corresponding artificial intelligence algorithm modules such as the denoising GAN module, super-resolution reconstruction module, color correction module, etc. in claim 1 to perform targeted quality restoration and optimization processing on the image, and then output the processed image to the post-processing module;

[0052] Post-processing and output module: Perform post-processing operations such as sharpening on the input restored and optimized image, and output the final processed scanned image as required.

[0053] The number of convolutional layers, pooling layers, and fully connected layers contained in the deep convolutional neural network in the feature extraction and analysis module, as well as the parameters of each layer, can be adjusted according to the characteristics of the scanned images in different application scenarios to achieve more accurate feature extraction and quality problem judgment.

[0054] Each artificial intelligence algorithm module in the repair and optimization module can be updated by updating training data, optimizing network structure or adjusting relevant algorithm parameters to adapt to the ever-changing scanning image quality issues and the performance improvement requirements brought about by the development of artificial intelligence technology, and ensure the system's continuous optimization of the scanning image quality processing effect.

[0055] Embodiment 1: Processing a document scanned image containing noise and poor definition;

[0056] Step 1: Image acquisition and preprocessing;

[0057] Use a conventional flatbed scanner to scan a paper document and obtain its color scanned image in JPEG format with a resolution of 300dpi. The color image is transferred to the preprocessing module and first grayscaled using the formula:

[0058] Gray=0.299×R+0.587×G+0.114×B

[0059] Convert the RGB value of each pixel of the image to grayscale value and convert it into a grayscale image. Then, normalize the grayscale image and use the formula Make the pixel value in the range [0,1], where x min and x max By traversing the entire image pixel value, the minimum and maximum values ​​are determined, x is the original pixel value, x norm is the normalized pixel value. After preprocessing, the image is passed to the feature extraction and analysis module.

[0060] Step 2: Feature extraction and analysis;

[0061] In the feature extraction and analysis module, a pre-trained deep convolutional neural network (CNN) is used. The network structure includes 5 convolutional layers, 3 pooling layers, and 2 fully connected layers. The convolution kernel size of the convolutional layer is set to 3×3, the step size is 1, and the padding method is SAME (that is, the output feature map size is guaranteed to be the same as the input feature map size in this dimension, and padding is performed if it is insufficient). For example, in the first convolutional layer, the input is a normalized grayscale image with a channel number of 1 (single channel grayscale image). According to the convolution calculation formula:

[0062]

[0063] The local features of the image are extracted by sliding convolution operations on the image with the convolution kernel, where each parameter is calculated according to the weights and biases determined during network initialization and training. The pooling layer uses a 2×2 maximum pooling window according to the formula:

[0064]

[0065] Downsampling is performed to gradually reduce the amount of data while retaining key features. After multiple layers of convolution and pooling, the fully connected layer integrates the extracted features and determines whether the image has noise interference and insufficient clarity quality problems, and passes the relevant feature information to the repair optimization module.

[0066] Step 3: Image quality repair and optimization;

[0067] Noise removal: The repair optimization module calls the trained generative adversarial network (GAN) to solve the noise problem. The generator network G structure is constructed with multiple transposed convolution layers, and its input is a random noise vector. Through adversarial training, the objective function is used:

[0068]

[0069] The generator is continuously optimized, where the data distribution of real noise-free images is p data (x) The distribution of random noise input to the generator is obtained by collecting a large number of high-quality noise-free document image samples. z (z) Using the standard normal distribution, during the training process, the discriminator D continuously distinguishes between real images and images generated by the generator, prompting the generator G to generate images that are closer to the real noise-free, thereby removing the noise in the scanned image.

[0070] Clarity improvement: To solve the problem of insufficient clarity, a super-resolution reconstruction algorithm based on deep learning is used. The super-resolution network used contains 8 residual blocks and the loss function L total =α×MSE+(1-α)×PerceptualLoss, the balance coefficient α is set to 0.8, mean square error (MSE) They are calculated according to the corresponding formulas, where the feature extraction function F refers to some layers of the pre-trained VGG network, N is the number of training samples (a large number of low-resolution and corresponding high-resolution document image pairs are used as samples when training the super-resolution network), and the network is trained by minimizing the total loss function to improve image clarity and make it easier to read and archive.

[0071] Step 4: Post-processing and output;

[0072] The image after noise removal and clarity enhancement is transmitted to the post-processing and output module, which performs edge sharpening operations on the image, using a common gradient-based sharpening algorithm to enhance the edge contours of image text and graphics to make the visual effect clearer. Finally, the processed image is output in the original JPEG format and saved in the designated document digital archive folder for subsequent retrieval and viewing.

[0073] The use of deep convolutional neural networks (CNN) can accurately determine whether images have noise interference and lack of clarity. Compared with ordinary processing methods that rely on manual experience or simple preset rules to judge problems, it is more objective and accurate, and can adapt to complex image situations caused by different document contents and different scanning conditions.

[0074] A generative adversarial network (GAN) is used to address noise, and adversarial training is used to generate images that are close to real noise-free. For insufficient clarity, a super-resolution reconstruction algorithm that combines mean square error and perceptual loss is used to improve clarity at the pixel level and perceptual level. Ordinary methods can only use general filtering denoising, simple interpolation and other means, which have limited effects and are prone to cause problems such as loss of image details or excessive smoothing. The targeted algorithm combination of this embodiment can better restore the clear and readable state of the document to meet archiving and viewing needs.

[0075] Embodiment 2: Processing scanned images of calligraphy and painting with color deviation;

[0076] Step 1: Image acquisition and preprocessing;

[0077] A high-precision scanner for calligraphy and painting is used to scan an ancient calligraphy and painting work, and a color scanned image with a resolution of 600dpi is obtained in TIFF format. In the preprocessing module, since the subsequent color correction needs to retain the original color information, it is not grayed out, but normalized. The pixel values ​​are normalized to the range [0,1] to ensure that the data meets the requirements of subsequent deep neural network operations, and then the image is passed to the feature extraction and analysis module.

[0078] Step 2: Feature extraction and analysis;

[0079] When the deep convolutional neural network (CNN) in the feature extraction and analysis module processes the scanned image of this painting and calligraphy, the convolution kernel size of the convolution layer is set to 5×5, the step size is 2 (the step size is appropriately increased to speed up feature extraction and adapt to the larger size characteristics of the painting and calligraphy images), and the filling mode is VALID (that is, no additional padding is performed, and the output feature map size is correspondingly reduced according to the convolution operation rules). After feature extraction and analysis of multiple convolutional layers and pooling layers (the network structure is different from Example 1, and has been specifically adjusted according to the requirements of feature recognition of painting and calligraphy images, including 6 convolutional layers, 4 pooling layers and 3 fully connected layers), it is judged that the image has a more obvious color deviation problem, such as the overall yellowish color tone, and the relevant feature information is passed to the restoration optimization module.

[0080] Step 3: Image quality repair and optimization;

[0081] The restoration and optimization module starts a color correction model based on a convolutional neural network. During training, the model uses a large number of standard color images of calligraphy and painting from different ages and under different preservation conditions, as well as corresponding scanned images with color deviations, as training samples to learn the color transformation rules between them. In the color correction process, the color difference formula in the CIELAB color space is used The loss function is used to measure the color difference between the corrected image and the standard reference image. By continuously adjusting the network parameters, the color of the output image is made as close as possible to the standard color provided by professional calligraphy and painting appraisal institutions, restoring the original color style of the calligraphy and painting works, which is beneficial to subsequent research, exhibitions and other applications.

[0082] Step 4: Post-processing and output;

[0083] The color-corrected scanned calligraphy and painting images are post-processed by fine-tuning the contrast and brightness (using conventional linear transformation methods to adjust the image pixel values) to achieve better visual presentation. Finally, the processed images are output in TIFF format and stored in a special calligraphy and painting digital resource library for art researchers and enthusiasts to view and appreciate.

[0084] By specially constructing a color correction model based on convolutional neural networks, we learn the color transformation rules between a large number of standard color images of calligraphy and painting and images with deviations, and use the CIELAB color space color difference formula to measure the color difference for precise correction. Compared with ordinary scanned image processing that may only rely on simple manual adjustments such as white balance and color balance or preset general correction parameters, it can better meet the characteristics of works such as calligraphy and painting that have extremely high requirements for color restoration, accurately restore the original color style of calligraphy and painting, and is conducive to applications such as art research and exhibitions.

[0085] The deep convolutional neural network used can adjust parameters such as the convolution layer and the pooling layer according to the requirements of calligraphy and painting image feature recognition, such as changing the convolution kernel size and step size. Compared with ordinary fixed parameters or general network structures, it is more adaptable when extracting calligraphy and painting image features, and can more keenly capture key feature information such as color deviation, providing accurate basis for subsequent correction.

[0086] Embodiment 3: Processing various quality issues existing in medical image scans;

[0087] Step 1: Image acquisition and preprocessing;

[0088] The patient's tomographic image is obtained through a medical CT scanner. The image format is DICOM (Digital Imaging and Communications in Medicine), the image is a grayscale image, and the resolution is 512×512. After it is passed to the preprocessing module, it is directly normalized according to the formula Ensure that the pixel value is in the range of [0,1] to facilitate subsequent neural network operations and then pass it to the feature extraction and analysis module.

[0089] Step 2: Feature extraction and analysis;

[0090] According to the characteristics of medical images, a specially designed deep convolutional neural network (CNN) is used for feature extraction and analysis. The network structure contains 8 convolutional layers, 5 pooling layers and 3 fully connected layers. The convolution kernel size is 3×3, the step size is 1, and the padding method is SAME. In order to better capture the subtle tissue structure features in medical images, the convolution kernel parameters of the convolution layer are optimized and adjusted by a large amount of medical image annotation data during training. After multiple layers of operation, it is analyzed that the image not only has certain noise interference, which affects the clarity of the tissue structure observation, but also has quality problems such as image blur and poor contrast in some areas due to problems such as scanning angles. These feature analysis results are passed to the repair optimization module.

[0091] Step 3: Image quality repair and optimization;

[0092] Noise removal: Generative adversarial networks (GANs) are used to remove noise from medical images. The generator network G structure is customized according to the characteristics of medical image noise. During the training process, the data distribution of real noise-free medical images is p data (x) is obtained from statistics of high-quality samples in a professional medical imaging database, and the random noise distribution p z (z) also uses the standard normal distribution, through the objective function:

[0093]

[0094] Optimize the generator so that the generated denoised images can restore the real tissue structure as much as possible and reduce the interference of noise on doctors' diagnosis.

[0095] Improve clarity and contrast: Use an algorithm that combines super-resolution reconstruction with contrast enhancement to improve image clarity and contrast. The super-resolution network incorporates prior knowledge of medical imaging when it is constructed, and the loss function comprehensively considers the differences at the pixel level and the perception level for optimization, enhancing the contrast between different tissue structures while improving the resolution. A contrast enhancement method based on histogram equalization is used to adjust the grayscale distribution of image pixels, making the originally blurred tissue boundaries in the image more clearly visible, making it easier for doctors to accurately judge the condition of the lesion.

[0096] Step 4: Post-processing and output;

[0097] After quality repair and optimization, the medical images are annotated in the subsequent post-processing and output modules (for example, key tissue structures, suspected lesion areas, and other text descriptions are annotated by superimposing text on the image), which makes it easier for doctors to quickly obtain key information when viewing. Finally, the processed medical images are output in DICOM format and stored in the hospital's imaging diagnosis system for doctors to use for diagnosis and analysis.

[0098] Customizing the deep convolutional neural network structure and parameters specifically for the characteristics of medical images can accurately analyze a variety of quality problems such as noise, image blur, and poor contrast. It is difficult for ordinary scanned image processing methods to simultaneously take into account the subtle characteristics of tissue structures in medical images and complex quality problems. At the same time, in the restoration and optimization stage, the algorithm combining the generative adversarial network denoising, super-resolution and contrast enhancement that conforms to the prior knowledge of medical images is used to better restore the morphology of tissue structures and enhance the contrast of key areas, making it easier for doctors to accurately diagnose lesions. This is an effect that is difficult to achieve with ordinary general-purpose image processing, and ordinary methods may not be able to effectively highlight important diagnostic information in medical images.

[0099] In the post-processing stage, annotation information is added to mark out key tissue structures, suspected lesion areas and other text descriptions to help doctors quickly obtain key information. Ordinary scanning image processing often only focuses on the image quality itself and lacks this auxiliary viewing function for medical application scenarios.

[0100] It can be seen from Example 1, Example 2 and Example 3 that the overall artificial intelligence technology, whether it is feature extraction and analysis, or repair optimization for different quality problems, relies on the powerful learning and pattern recognition capabilities of deep neural networks. Compared with ordinary scanning image processing that relies on fixed algorithms, manual experience, and preset parameters, it can more accurately deal with various complex quality problems that appear in scanning images of different types and application scenarios, and has strong adaptability and versatility.

[0101] In addition, the image quality can be improved from multiple angles, including noise removal, clarity improvement, color correction, contrast enhancement, and adding auxiliary annotations to improve the quality of scanned images in all dimensions. Ordinary processing methods can often only solve problems in a single or a few aspects, and it is difficult to meet today's diverse and high-quality application scenarios, such as medical diagnosis, preservation of artworks, and large-scale digital archiving of documents.

[0102] Furthermore, each module in the system (such as feature extraction network, repair optimization algorithm module, etc.) is scalable. With the continuous development of artificial intelligence technology and the accumulation of more high-quality training data, the network structure can be easily updated, the algorithm parameters can be optimized, and the processing effect can be continuously improved. However, it is relatively difficult to update and improve ordinary scanning image processing methods, and it is difficult to keep up with the growing quality requirements of technological development and practical applications.

[0103] It should be noted that in the comparative embodiment, the comparative embodiment adopts traditional processing methods to process images in various situations, and relies more on traditional, preset fixed parameters and relatively simple algorithms to deal with quality problems existing in scanned images. Compared with the previously introduced artificial intelligence-based processing method, there are often certain limitations in terms of the accuracy of processing effects, adaptability to complex situations, and the degree of quality improvement.

[0104] Table 1: Comparison of experimental data between Example 1 and Comparative Example

[0105]

[0106] Table 2: Comparison of experimental data between Example 2 and Comparative Example

[0107]

[0108] Table 3: Comparison of experimental data between Example 3 and Comparative Example

[0109]

[0110] It can be clearly seen from the comparison of the above tables that in different application scenarios, the scanning image processing method based on artificial intelligence has better performance in various measurement indicators than the ordinary processing method, and can more effectively improve the quality of the scanned image and the processing efficiency, etc., reflecting the advantages of the present invention in the field of scanning image processing.

[0111] Specifically, in each embodiment, key code examples and corresponding analysis are written in Python language.

[0112] Furthermore, if Figure 1 As shown, its function is to convert a color image into a grayscale image. According to the weighting coefficients determined by the human eye's sensitivity to different colors, the red, green, and blue channel values ​​of the color image are calculated to obtain the corresponding grayscale values. Finally, the data type is converted to ensure that the format meets the requirements of subsequent processing, so that the image can use a single channel to represent the brightness and darkness, simplifying the data dimension for subsequent processing.

[0113] Furthermore, if Figure 2 As shown in the figure, the purpose is to map the pixel values ​​of the input image (grayscale or color) to the interval [0,1]. By distinguishing the image type (grayscale or color), the minimum and maximum values ​​are calculated respectively to complete the normalization. This can unify the data level, making it easier for subsequent neural network algorithms to process image data more stably and efficiently, thereby improving the processing effect.

[0114] Furthermore, if Figure 3As shown in the figure, the core operation of the convolution layer in the convolutional neural network is simulated. The output feature map size is calculated according to the convolution kernel size and the input image size. Then, through multiple layers of loop traversal, according to the convolution operation rules, the convolution kernel is used to perform weighted summation on the local area of ​​the input image and add a bias term to extract local features of the image at different scales, and generate an output feature map for further processing in subsequent network layers. This is a key step in extracting image features in deep convolutional neural networks.

[0115] Furthermore, if Figure 4-6 As shown in the figure, as a part of GAN, it receives random noise vector input, and gradually maps the input dimension to the appropriate output dimension through a combination of multiple layers of fully connected layers and corresponding activation functions (ReLU, tanh, etc.), and uses the nonlinear characteristics of the activation function to enhance the network expression ability, and finally generates data that meets certain requirements (such as denoised images, etc.), with the purpose of generating realistic data that can deceive the discriminator. It is also a key component of GAN, and its input is data that may be real or generated. Through the processing of multiple layers of fully connected layers and activation functions (ReLU, sigmoid, etc.), the input data is mapped to a value representing the probability of it being real data, so as to judge the source of the input data, aiming to accurately distinguish between real data and data generated by the generator, and alternately train the discriminator and generator in each training cycle. The discriminator first learns to distinguish between real images and fake images generated by the generator, and the generator adjusts its own parameters according to the judgment results of the discriminator to generate more realistic images. Through multiple rounds of iterations, the two are constantly competing, so that the images generated by the generator are closer and closer to the real images. However, the example is just a simple illustration, and many details such as loss calculation and optimizer need to be improved in actual applications.

[0116] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A scanning image processing method based on artificial intelligence, characterized in that: The following steps are involved: S1. Image acquisition and preprocessing: First, obtain the scanned image through a conventional scanning device, convert it into a common digital image format, and perform basic preprocessing operations. This includes but is not limited to grayscale processing of the image and converting the RGB value of the color image into grayscale value; S2. Feature extraction and analysis: A deep convolutional neural network is used to extract features of the scanned image. The network structure includes multiple convolutional layers, pooling layers, and fully connected layers. The convolutional layer performs convolution operations by sliding the convolution kernel on the image to extract local features of different scales. After multiple layers of convolution and pooling operations, the extracted features are integrated and mapped to a specific feature space through the fully connected layer, which is used for subsequent classification or regression tasks to determine the type and degree of quality problems in the image. S3. Image quality repair and optimization: Noise removal: If noise is detected in the image, a generative adversarial network is used to generate a denoised image. Clarity improvement: For situations where clarity is insufficient, a super-resolution reconstruction algorithm based on deep learning is used to improve image clarity by learning the mapping relationship between low-resolution images and corresponding high-resolution images; Color correction: Use deep neural networks to learn the color transformation rules under different lighting conditions, scanning equipment and other factors, and establish a mapping relationship between the input scanned image color and the standard color space. A color correction model based on convolutional neural network is adopted, and its training process optimizes the network parameters by minimizing the color difference loss; Post-processing and output: Post-process the quality-restored and optimized images, such as sharpening edges to further enhance the image visual effect, and then output the processed images in the specified format for subsequent storage, viewing or other applications.

2. The method for processing scanned images based on artificial intelligence according to claim 1, characterized in that: In step S2, the convolution kernel size of the convolution layer of the deep convolutional neural network is set to 3×3 or 5×5, the step size is 1 or 2, and the filling method is SAME or VALID. Different setting combinations are used to adapt to the feature extraction requirements of different types of scanned images.

3. The method and system for processing scanned images based on artificial intelligence according to claim 1, characterized in that: In step S3, the super-resolution network used by the super-resolution reconstruction algorithm includes multiple residual blocks, and the network's learning and reconstruction capabilities for image features are improved through residual learning, so as to better improve the clarity of the scanned image.

4. The method and system for processing scanned images based on artificial intelligence according to claim 1, characterized in that: In step S3, the training sample acquisition method of the color correction model based on the convolutional neural network is to collect standard color images of the target object in different ages and under different preservation conditions and the corresponding scanned images with color deviations, and optimize the network parameters by learning the color transformation rules between the two to achieve accurate correction of the color of the scanned image.

5. A scanning image processing system based on artificial intelligence, characterized in that: include: Image acquisition module: used to connect to the scanning device, obtain the scanned original image, and transmit it to the subsequent module for processing; Preprocessing module: receiving the image from the image acquisition module, performing preprocessing operations on the image according to the grayscale, normalization and other methods described in claim 1, and outputting the preprocessed image to the feature extraction module; Feature extraction and analysis module: Built-in deep convolutional neural network and other related artificial intelligence algorithms to extract and analyze the features of the input pre-processed image, determine the type and degree of quality problems in the image, and pass the analysis results to the repair optimization module; Restoration and optimization module: based on the received feature analysis results, call the corresponding artificial intelligence algorithm modules such as the denoising GAN module, super-resolution reconstruction module, color correction module, etc. as described in claim 1 to perform targeted quality restoration and optimization processing on the image, and then output the processed image to the post-processing module; Post-processing and output module: Perform post-processing operations such as sharpening on the input restored and optimized image, and output the final processed scanned image as required.

6. The scanning image processing system based on artificial intelligence according to claim 5, characterized in that: The number of convolutional layers, pooling layers, and fully connected layers included in the deep convolutional neural network in the feature extraction and analysis module, as well as the parameters of each layer, can be adjusted according to the characteristics of the scanned images in different application scenarios to achieve more accurate feature extraction and quality problem judgment.

7. The scanning image processing system based on artificial intelligence according to claim 5, characterized in that: Each artificial intelligence algorithm module in the repair optimization module can be updated by updating training data, optimizing network structure or adjusting relevant algorithm parameters to adapt to the ever-changing scanning image quality problems and the performance improvement requirements brought about by the development of artificial intelligence technology, thereby ensuring the system's continuous optimization of the scanning image quality processing effect.

Images