Dental fluorosis visual detection image enhancement method

Abstract

The invention provides a dental fluorosis visual detection image enhancement method, and relates to the technical field of image enhancement. The invention provides a dental fluorosis image data processing flow. The dental fluorosis image data processing flow comprises the steps of obtaining a dental fluorosis image data set, constructing a preliminary enhancement module, constructing a low-high frequency feature interaction module, constructing a deepening enhancement module and constructing a dental fluorosis image enhancement model. Wherein the preliminary enhancement module decomposes a dental fluorosis image into a low-frequency component and a high-frequency component through discrete wavelet transform and preliminarily enhances the dental fluorosis image in combination with discrete cosine transform and an attention mechanism, and the low-frequency and high-frequency feature interaction module is used for interacting features output by a low-frequency branch and a high-frequency branch in the preliminary enhancement module. And the deepening enhancement module utilizes the generative adversarial network to enhance the low-frequency and high-frequency features after interaction of the low-frequency and high-frequency feature interaction module, and fuses the low-frequency and high-frequency features to output a dental fluorosis enhanced image.

Description

Technical Field

[0001] The present invention belongs to the technical field of image enhancement, and in particular relates to an image enhancement method for visual detection of dental fluorosis. Background Art

[0002] Dental fluorosis is a dental disease caused by excessive intake of fluoride, which usually manifests as white or brown spots on the surface of the teeth. In severe cases, problems such as reduced hardness and structural damage of the teeth may occur. Early diagnosis of dental fluorosis is of great significance for prevention and treatment. Currently, the diagnosis of dental fluorosis mainly relies on dentists to observe the appearance of the teeth with the naked eye, or to perform auxiliary examinations through X-ray images. However, these diagnostic methods have certain limitations.

[0003] With the development of computer vision and deep learning technologies, image processing technology has been widely used in the field of medical imaging. In particular, deep learning methods have made significant progress in image feature extraction and enhancement. However, feature enhancement methods for fluorosis images are still relatively lacking. Existing research mostly focuses on ordinary tooth images or other types of medical image processing. There are still relatively large challenges in the detection of fluorosis lesions.

[0004] Traditional deep learning methods often face problems such as excessive image noise, loss of detail information, and insufficient image contrast when processing fluorosis images. In addition, standard convolutional neural networks may not be able to effectively extract the detailed features of fluorosis images. Existing methods generally rely on end-to-end learning and ignore the efficient use of image frequency domain information, which may lead to poor fluorosis image enhancement. Therefore, a fluorosis image enhancement method combining discrete wavelet transform, attention mechanism and generative adversarial network is proposed. Summary of the Invention

[0005] The present invention provides an image enhancement method for visual detection of dental fluorosis, aiming to propose a dental fluorosis image enhancement model, wherein a preliminary enhancement module uses discrete wavelet transform to decompose the dental fluorosis image into low-frequency components and high-frequency components, extracts features of the low-frequency branch through multi-scale convolution and then fuses them, then converts them to the frequency domain through discrete cosine transform to process the amplitude and phase, and finally inversely transforms them back to the spatial domain and fuses them with the original spatial domain features to obtain low-frequency embedded features, and obtains high-frequency embedded features of the high-frequency branch through an attention mechanism and a feedforward neural network; wherein the low-frequency and high-frequency feature interaction module realizes the interaction between low-frequency and high-frequency features through linear projection and attention mechanism, so that the low-frequency features obtain local detail information from the high-frequency features, and the high-frequency features obtain global context information from the low-frequency features; wherein the deepening enhancement module uses a generative adversarial network combined with a noise vector to generate the restored frequency components, and finally uses an inverse discrete wavelet transform to generate a dental fluorosis enhanced image.

[0006] The present invention aims to propose a dental fluorosis image enhancement model and provide a dental fluorosis visual detection image enhancement method, which includes the following steps.

[0007] S1. Acquire a dental fluorosis image dataset, and preprocess the dental fluorosis image dataset. The preprocessing operations include denoising, normalization, and image segmentation.

[0008] S2. Construct a preliminary enhancement module, which uses discrete wavelet transform to decompose the fluorosis image into low-frequency components and high-frequency components, combines discrete cosine transform and attention mechanism to preliminarily enhance the fluorosis image, designs adaptive fusion parameters based on feature similarity, and fuses low-frequency features according to the similarity of spatial domain embedded features and frequency domain embedded features.

[0009] S3. Construct a low- and high-frequency feature interaction module, which is used to interact the features output by the low-frequency branch and the high-frequency branch in the preliminary enhancement module to achieve cross-frequency mutual enhancement.

[0010] S4. Construct a deepening enhancement module, which uses an improved generative adversarial network to enhance the low and high frequency features after the interaction of the low and high frequency feature interaction module, and fuses the enhanced low and high frequency features to generate a fluorosis enhanced image. By optimizing the discriminator's adversarial loss, perceptual loss and feature consistency loss, the generator is guided to generate a high-quality fluorosis enhanced image.

[0011] S5. Construct a dental fluorosis image enhancement model, wherein the dental fluorosis image enhancement model consists of a preliminary enhancement module, a low- and high-frequency feature interaction module, and a deep enhancement module.

[0012] Preferably, in step S1, for obtaining a dental fluorosis image dataset, dental fluorosis images taken under different lighting conditions such as natural light, LED light source, X-ray, and from multiple angles such as front, side, and top view are collected, including images of mild, moderate, and severe degrees of dental fluorosis. Denoising, normalization, and image segmentation preprocessing operations are performed on the dental fluorosis image dataset, and the preprocessed dental fluorosis image dataset is divided into a training set and a test set.

[0013] Preferably, in step S2, the adaptive fusion parameters based on feature similarity are designed, and the specific steps are as follows: Spatial domain embedding features in low frequency components of dental fluorosis images and frequency domain embedding features , , ,calculate The mean and variance ,calculate The mean and variance ,according to and The similarity in different statistical characteristics is used to calculate the multi-dimensional similarity vector S. The specific formula is: ; Where, is the covariance; Design of adaptive fusion parameters based on multi-dimensional similarity vector S , the specific formula is: ; Where w is the learnable weight and b is the learnable bias. w and b are updated by gradient descent during training.

[0014] Preferably, in step S2, for the adaptive fusion parameters based on feature similarity, the embedding features of the low-frequency branch are fused by calculating the similarity of the spatial domain embedding features and the frequency domain embedding features, so that the embedding features of the fused low-frequency branch can more accurately reflect the global information of the fluorosis image, and the adaptive fusion parameters are dynamically adjusted according to different features, so that the fluorosis image is clearer and over-smoothing or loss of details is avoided.

[0015] Preferably, in step S2, the method for constructing the preliminary enhancement module is: S21. Input dental fluorosis image , , 、 and are the height, width and number of channels of the dental fluorosis image respectively. First, the dental fluorosis image is transformed into The fluorosis image is decomposed into multiple frequency sub-bands and the low-pass and high-pass filters of Daubechies wavelet are used to filter the fluorosis image. Transform to obtain the low-frequency subband , horizontal high frequency sub-band , vertical high frequency sub-band and diagonal high frequency subbands ,in ,The Daubechies wavelet can better extract multi-scale features, and is ,particularly suitable for detecting dental fluorosis images with obvious ,detail changes; S22, the preliminary enhancement module consists of a low-frequency branch and a high-frequency branch. For the low-frequency branch, the convolution kernel is used. 、 and Convolution extracts low-frequency components The multi-scale features of , The scale is The convolution kernel, Is the convolution operation, and then through the convolution kernel is The convolution of different scales can enhance the detection capability of spots of different sizes and obtain the spatial domain embedding feature. , , 、 and is the weight coefficient. During the training process, the weight coefficient of the convolution kernel is updated through back propagation and gradient descent. The fluorosis image features in the spatial domain are transformed by discrete cosine transform. Convert to frequency domain features , , is the discrete cosine transform, and is the coordinate in the frequency domain, and the frequency domain features are calculated Amplitude , , and plural The real and imaginary parts of Phase , , the separation optimization of amplitude and phase can enhance the overall contrast and texture details of the image respectively, and then the amplitude of the frequency domain features is and phase Through two convolution kernels respectively The convolution is optimized and phase , use the inverse discrete cosine transform to transform the optimized amplitude and phase Convert back to image space to get frequency domain embedding features , , is the inverse discrete cosine transform, and finally uses the adaptive fusion parameter The spatial domain embedding features and the frequency domain embedding features are fused to obtain the embedding features of the low-frequency branch , ; S23. For high-frequency branches, high-frequency components are embedded through convolution projection operations. , , , the embedded features are input into the attention mechanism, and the attention mechanism is applied to the high-frequency feature processing of dental fluorosis. It can dynamically focus on the key high-frequency features in the dental fluorosis image and divide them into query Q, key K, value V and additional features L. , calculate the attention weight , , Is the dimension of the key feature, calculate the weighted feature , , and finally the weighted features Input into the feedforward neural network to obtain the embedded features of the high-frequency branches , .

[0016] Preferably, in step S2, for the preliminary enhancement module, the use of discrete wavelet transform can effectively separate the different frequency components of the fluorosis image, which is conducive to retaining the details and overall information of the image; for the low-frequency branch, the multi-scale convolution operation extracts features of different scales, which is conducive to capturing different pathological features in the fluorosis image, and the discrete cosine transform converts the spatial domain features to the frequency domain to process the amplitude and phase, which can optimize the features from the frequency domain perspective and enhance the expressive ability of the features; for the high-frequency branch, the attention mechanism can enable the model to focus on the key areas in the high-frequency information, especially the pathological areas; the preliminary enhancement module is a key link in the entire fluorosis image enhancement process. This module fully combines the characteristics of the frequency domain and the spatial domain, and realizes the efficient separation and optimization of different frequency information of the fluorosis image through discrete wavelet transform, discrete cosine transform, multi-scale convolution operation and attention mechanism.

[0017] Preferably, in step S3, the method for constructing the low- and high-frequency feature interaction module is: Input the low-frequency embedding features output by the preliminary enhancement module and high-frequency embedded features , , , first embed the high-frequency features Perform a segmentation operation to obtain the high-frequency horizontal component , vertical component and the diagonal components , and then add the coefficients of the three components to get the aggregated high-frequency embedding features , , , use linear projection to embed low-frequency features and aggregated high-frequency embedded features Mapping to the attention space, the attention mechanism is applied to the interactive processing of low-frequency and high-frequency features, and the correlation between features is dynamically captured through query, key and value. In the fluorosis detection task, the attention mechanism can effectively focus on the key feature area while suppressing irrelevant or redundant information, and obtain the query and key , , , get the low frequency value and high frequency values , , , calculate the attention weight , is the dimension of the key feature, and the weighted summation is used to obtain the low-frequency feature and high-frequency features , , .

[0018] Preferably, in step S3, for the low- and high-frequency feature interaction module, the interaction between low-frequency and high-frequency features is realized through linear projection and attention mechanism, so that the low-frequency features can obtain local detail information from the high-frequency features, and the high-frequency features can obtain global context information from the low-frequency features. For fluorosis images, the low-frequency features and high-frequency feature information complement each other, so that the model can understand the image more comprehensively and enhance the features of the lesion area of ​​the fluorosis image; the low- and high-frequency feature interaction module is designed to enhance the advantages of low-frequency features in global structural expression and the ability of high-frequency features in capturing details, and provide more comprehensive and accurate feature support for the enhancement of the lesion area of ​​fluorosis images, so that the model not only effectively retains the overall information and detail features of the fluorosis image, but also significantly improves the prominence and recognizability of the lesion area of ​​the fluorosis image.

[0019] Preferably, in step S4, the construction method of the deepening enhancement module is: S41, input low-frequency features output by low-frequency and high-frequency feature interaction module and high-frequency features , that is, the low-frequency component , , that is, the high-frequency component , , , calculate the real image and enhanced low-frequency components The residual between , , calculate the real image and enhanced high-frequency components The residual between , , the residual and As the target of the generator, the enhanced low-frequency component and the enhanced high-frequency components Input to the generator The generator can capture complex nonlinear features, improve the details of fluorosis images, and combine the noise vector and Generate restored frequency components and , , , the low-frequency component and high-frequency component generated by the inverse discrete wavelet transform fusion generator are used to obtain the enhanced fluorosis image , , IDWT is the inverse discrete wavelet transform; S42. The loss function of the discriminator introduces perceptual loss and feature consistency loss on the basis of adversarial loss to improve the effect of the generator in generating fluorosis enhanced images, where the adversarial loss is , , is the expected value, is the discriminator, and the perceptual loss is , , is a feature extractor, is the L2 norm, calculating the Euclidean distance between two features, and the feature consistency loss is , , is the L2 norm, which is used to measure the degree of change of features before and after enhancement. The overall loss of the discriminator is , , and is a weight coefficient that controls the importance of perceptual loss and feature consistency loss in the total loss.

[0020] Preferably, in step S4, for the deepening enhancement module, the residual is calculated as the target of the generator, so that the generated restored frequency component can more accurately improve the deficiencies in the fluorosis image enhancement, thereby generating an image that is more consistent with the real image characteristics and has a better enhancement effect. The generative adversarial network can improve the quality and authenticity of the generated image through the application of the generator and the discriminator, and help to generate clearer images with significantly enhanced lesion features. The discriminator guides the generator through the loss function to continuously improve the enhancement effect of the fluorosis image; the deepening enhancement module aims to enable the generator to generate high-quality fluorosis enhanced images, and the discriminator continuously promotes the generator to improve the quality and authenticity of the fluorosis image through the guidance of the discriminant loss.

[0021] Compared with the prior art, the present invention has the following technical effects: The technical solution provided by the present invention proposes a fluorosis image enhancement model, wherein the preliminary enhancement module uses discrete wavelet transform to decompose the fluorosis image into low-frequency components and high-frequency components, extracts features from the low-frequency branch through multi-scale convolution and then fuses them, then converts them to the frequency domain through discrete cosine transform to process the amplitude and phase, and finally inversely transforms them back to the spatial domain and fuses them with the original spatial domain features to obtain low-frequency embedded features. For the high-frequency branch, high-frequency embedded features are obtained through attention mechanism and feedforward neural network; wherein the low-frequency and high-frequency feature interaction module realizes the interaction between low-frequency and high-frequency features through linear projection and attention mechanism, so that the low-frequency features obtain local detail information from the high-frequency features, and the high-frequency features obtain global context information from the low-frequency features; wherein the deepening enhancement module uses generative adversarial network combined with noise vector to generate the restored frequency components, and finally uses inverse discrete wavelet transform to generate fluorosis enhanced image. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a flow chart of dental fluorosis image data processing provided by the present invention.

[0023] Figure 2 This is a structural diagram of the preliminary enhancement module provided by the present invention.

[0024] Figure 3 This is a structural diagram of the low- and high-frequency feature interaction module provided by the present invention.

[0025] Figure 4 This is the effect diagram of the dental fluorosis image before enhancement provided by the present invention.

[0026] Figure 5 This is the effect diagram of the enhanced dental fluorosis image provided by the present invention. DETAILED DESCRIPTION

[0027] The present invention provides an image enhancement method for visual detection of dental fluorosis, aiming to propose a dental fluorosis image enhancement model, wherein a preliminary enhancement module uses discrete wavelet transform to decompose the dental fluorosis image into low-frequency components and high-frequency components, extracts features of the low-frequency branch through multi-scale convolution and then fuses them, then converts them to the frequency domain through discrete cosine transform to process the amplitude and phase, and finally inversely transforms them back to the spatial domain and fuses them with the original spatial domain features to obtain low-frequency embedded features, and obtains high-frequency embedded features of the high-frequency branch through an attention mechanism and a feedforward neural network; wherein the low-frequency and high-frequency feature interaction module realizes the interaction between low-frequency and high-frequency features through linear projection and attention mechanism, so that the low-frequency features obtain local detail information from the high-frequency features, and the high-frequency features obtain global context information from the low-frequency features; wherein the deepening enhancement module uses a generative adversarial network combined with a noise vector to generate the restored frequency components, and finally uses an inverse discrete wavelet transform to generate a dental fluorosis enhanced image.

[0028] See Figure 1As shown, an image enhancement method for visual detection of dental fluorosis in an embodiment of the present application.

[0029] S1. Acquire a dental fluorosis image dataset, and perform preprocessing on the dental fluorosis image dataset. The preprocessing operations include denoising, normalization, and image segmentation.

[0030] Furthermore, in step S1, a fluorosis image dataset is obtained, and denoising is performed using a median filtering method. For fluorosis images, median filtering can effectively remove isolated noise points such as salt and pepper noise, while better preserving image edges and detail information; a linear normalization method is used to map the grayscale values ​​of the fluorosis images to a specific interval, so that the image data has the same scale, which is convenient for subsequent processing and analysis, while highlighting the useful information in the image; a threshold-based segmentation method is used to select a suitable threshold according to the grayscale difference between the region of interest and the background in the fluorosis image, and the pixels in the image are divided into two categories: the region of interest and the background, thereby achieving segmentation; the preprocessed fluorosis image dataset is divided into a training set and a test set.

[0031] S2. Construct a preliminary enhancement module, which uses discrete wavelet transform to decompose the fluorosis image into low-frequency components and high-frequency components, combines discrete cosine transform and attention mechanism to preliminarily enhance the fluorosis image, designs adaptive fusion parameters based on feature similarity, and fuses low-frequency features according to the similarity of spatial domain embedded features and frequency domain embedded features.

[0032] Furthermore, in step S2, adaptive fusion parameters based on feature similarity are designed, and the specific steps are as follows: Spatial domain embedding features in low frequency components of dental fluorosis images and frequency domain embedding features , , ,calculate The mean and variance , the specific calculation formula is: ; ; calculate The mean and variance , the specific calculation formula is: ; ; according to and The similarity in different statistical characteristics is used to calculate the multi-dimensional similarity vector S. The specific formula is: ; Where, is the covariance; Design of adaptive fusion parameters based on multi-dimensional similarity vector S , the value range during implementation is 0 to 1, and the specific formula is: ; Where w is the learnable weight, which is initially set to 0.1 during implementation and ranges from -0.1 to 0.1; b is the learnable bias, which is initially set to 0 during implementation.

[0033] Furthermore, in the step S2, the structure of the preliminary enhancement module is as follows: Figure 2 As shown, its construction method is: S21. Input dental fluorosis image , , 224, 224 and 3 are the height, width and number of channels of the dental fluorosis image respectively. First, the dental fluorosis image is transformed into The fluorosis image is decomposed into multiple frequency sub-bands and the low-pass and high-pass filters of Daubechies wavelet are used to filter the fluorosis image. Transform to obtain the low-frequency subband , horizontal high frequency sub-band , vertical high frequency sub-band and diagonal high frequency subbands ,in ; S22, the preliminary enhancement module consists of a low-frequency branch and a high-frequency branch. For the low-frequency branch, the convolution kernel is used. 、 and Convolution extracts low-frequency components The multi-scale features of , The scale is The convolution kernel, Is the convolution operation, and then through the convolution kernel is Convolution fuses features of different scales to obtain spatial domain embedding features , , 、 and is the weight coefficient. The value range of the weight coefficient in the implementation process is -0.01 to 0.01. The fluorosis image features in the spatial domain are transformed by discrete cosine transform. Convert to frequency domain features , , is the discrete cosine transform, and is the coordinate in the frequency domain, and the frequency domain features are calculated Amplitude , , and plural The real and imaginary parts of Phase , , and then the amplitude of the frequency domain feature and phase Through two convolution kernels respectively The convolution is optimized and phase , use the inverse discrete cosine transform to transform the optimized amplitude and phase Convert back to image space to get frequency domain embedding features , , is the inverse discrete cosine transform, and finally uses the adaptive fusion parameter The spatial domain embedding features and the frequency domain embedding features are fused to obtain the embedding features of the low-frequency branch , ; S23. For high-frequency branches, high-frequency components are embedded through convolution projection operations. , , , input the embedded features into the attention mechanism and split them into query Q, key K, value V and additional features L, , calculate the attention weight , , the value range during implementation is 0 to 1, Is the dimension of the key feature, calculate the weighted feature , , and finally the weighted features Input into the feedforward neural network to obtain the embedded features of the high-frequency branches , .

[0034] S3. Construct a low- and high-frequency feature interaction module, which is used to interact with the features output by the low-frequency branch and the high-frequency branch in the preliminary enhancement module to achieve cross-frequency mutual enhancement.

[0035] Furthermore, in the step S3, the structure of the low- and high-frequency feature interaction module is as follows: Figure 3 As shown, its construction method is: Input the low-frequency embedding features output by the preliminary enhancement module and high-frequency embedded features , , , first embed the high-frequency features Perform a segmentation operation to obtain the high-frequency horizontal component , vertical component and the diagonal components , and then add the coefficients of the three components to get the aggregated high-frequency embedding features , , , use linear projection to embed low-frequency features and aggregated high-frequency embedded features Mapped to the attention space, we get the query and key , , , get the low frequency value and high frequency values , , , calculate the attention weight , is the dimension of the key feature, and the weighted summation is used to obtain the low-frequency feature and high-frequency features , , .

[0036] S4. Construct a deepening enhancement module, which uses an improved generative adversarial network to enhance the low and high frequency features after the interaction of the low and high frequency feature interaction module, and fuses the enhanced low and high frequency features to generate a fluorosis enhanced image. By optimizing the discriminator's adversarial loss, perceptual loss and feature consistency loss, the generator is guided to generate a high-quality fluorosis enhanced image.

[0037] Furthermore, in step S4, the construction method of the deepening enhancement module is: S41, input low-frequency features output by low-frequency and high-frequency feature interaction module and high-frequency features , that is, the low-frequency component , , that is, the high-frequency component , , , calculate the real image and enhanced low-frequency components The residual between , , calculate the real image and enhanced high-frequency components The residual between , , the residual and As the target of the generator, the enhanced low-frequency component and the enhanced high-frequency components Input to the generator In the noise vector and Generate restored frequency components and , , , the low-frequency component and high-frequency component generated by the inverse discrete wavelet transform fusion generator are used to obtain the enhanced fluorosis image , , IDWT is the inverse discrete wavelet transform; S42. The loss function of the discriminator introduces perceptual loss and feature consistency loss on the basis of adversarial loss, where adversarial loss is , , is the expected value, is the discriminator, and the perceptual loss is , , is a feature extractor, is the L2 norm, calculating the Euclidean distance between two features, and the feature consistency loss is , , is the L2 norm, which is used to measure the degree of change of features before and after enhancement. The overall loss of the discriminator is , , and Is the weight coefficient, which controls the importance of perceptual loss and feature consistency loss in the total loss. During the implementation, The initial value is set to 0.2, and the range is 0.05 to 0.5. The initial value is set to 0.5 and the range is 0.1 to 1.

[0038] S5. Construct a dental fluorosis image enhancement model, wherein the dental fluorosis image enhancement model consists of a preliminary enhancement module, a low- and high-frequency feature interaction module, and a deep enhancement module.

[0039] Furthermore, in step S5, for the fluorosis image enhancement model, the fluorosis image training set is input into the model in batches, the output of the model is calculated according to the forward propagation, and then the loss value is calculated according to the loss function. The gradient of the loss with respect to the model parameters is calculated through back propagation, and finally the model parameters are updated using the optimization algorithm. During the training process, the performance of the model is regularly evaluated on the fluorosis image validation set, and indicators such as the peak signal-to-noise ratio and the structural similarity index are monitored to determine whether the model is overfitting or underfitting. After the training is completed, the final model is evaluated using the fluorosis image test set to compare the quality and visual effects of the images before and after enhancement, and to analyze the effectiveness and accuracy of the model in the fluorosis image enhancement task.

[0040] Furthermore, in step S5, for the dental fluorosis image enhancement model, it is written in Python language and uses the Pytorch framework. The batch size is 32, the learning rate is set to 0.0001, the training rounds are 100, and the generator in the improved generative adversarial network uses the Adam optimizer and the discriminator uses the stochastic gradient descent optimizer. The dental fluorosis image preprocessed in step S1 is input into the dental fluorosis image enhancement model, as shown in Figure 5. Figure 4 As shown, Figure 4 The effect of the fluorosis image before enhancement is shown. The features of the fluorosis image are preliminarily enhanced through the preliminary enhancement module to obtain the preliminarily enhanced features. The preliminarily enhanced features are mutually enhanced across frequencies through the low-high frequency feature interaction module to obtain the low-high frequency features after interaction. The low-high frequency features after interaction are fused with the enhanced low-high frequency features through the deepening enhancement module to generate the fluorosis enhanced image. Figure 5 As shown, Figure 5 The image shows the effect of dental fluorosis enhancement after being processed by the dental fluorosis image enhancement model.

[0041] The above are only preferred embodiments of the present invention. It should be pointed out that those skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A method for enhancing dental fluorosis visual detection images, characterized in that: The following steps are involved: S1. Acquire a dental fluorosis image dataset and perform preprocessing on the dental fluorosis image dataset, wherein the preprocessing operations include denoising, normalization, and image segmentation; S2. Constructing a preliminary enhancement module. The module decomposes the dental fluorosis image into low-frequency and high-frequency components using discrete wavelet transform, combines discrete cosine transform and attention mechanism to preliminarily enhance the dental fluorosis image, designs adaptive fusion parameters based on feature similarity, and fuses low-frequency features based on the similarity of spatial domain embedded features and frequency domain embedded features. S3. Construct a low- and high-frequency feature interaction module, wherein the low- and high-frequency feature interaction module is used to interact with the features output by the low-frequency branch and the high-frequency branch in the preliminary enhancement module to achieve cross-frequency mutual enhancement; S4. Constructing a deepening enhancement module, wherein the deepening enhancement module utilizes an improved generative adversarial network to enhance the low- and high-frequency features after interaction in the low- and high-frequency feature interaction module, fuses the enhanced low- and high-frequency features to generate a fluorosis-enhanced image, and guides the generator to generate a high-quality fluorosis-enhanced image by optimizing the adversarial loss, perceptual loss, and feature consistency loss of the discriminator; S5. Construct a dental fluorosis image enhancement model, wherein the dental fluorosis image enhancement model consists of a preliminary enhancement module, a low- and high-frequency feature interaction module, and a deep enhancement module.

2. The method for visual detection of dental fluorosis according to claim 1, characterized in that: In step S2, adaptive fusion parameters based on feature similarity are designed. The specific steps are as follows: Spatial domain embedding features in low frequency components of dental fluorosis images and frequency domain embedding features , , ,calculate The mean and variance ,calculate The mean and variance ,according to and The similarity in different statistical characteristics is used to calculate the multi-dimensional similarity vector S. The specific formula is: ； Where, is the covariance; Design of adaptive fusion parameters based on multi-dimensional similarity vector S , the specific formula is: ； Where w is the learnable weight and b is the learnable bias.

3. The method for visual detection of dental fluorosis according to claim 2, characterized in that: In step S2, the method for constructing the preliminary enhancement module is as follows: S21. Input a dental fluorosis image. First, use discrete wavelet transform to decompose the dental fluorosis image into multiple frequency sub-bands. Then, transform the dental fluorosis image using Daubechies wavelet low-pass and high-pass filters to obtain low-frequency sub-bands, horizontal high-frequency sub-bands, vertical high-frequency sub-bands, and diagonal high-frequency sub-bands. S22. The preliminary enhancement module consists of a low-frequency branch and a high-frequency branch. For the low-frequency branch, convolutions of different sizes are used to extract multi-scale features of the low-frequency component, and then the features of different scales are fused by convolution with a convolution kernel of 1×1 to obtain spatial domain embedding features. The spatial domain embedding features are converted into frequency domain features by discrete cosine transform, the amplitude of the frequency domain features is calculated, the phase of the frequency domain features is calculated, and then the amplitude and phase of the frequency domain features are respectively obtained by two convolution kernels of 1×1 to obtain the optimized amplitude and phase. The optimized amplitude and phase are converted back to the image space using inverse discrete cosine transform to obtain frequency domain embedding features, and finally the adaptive fusion parameters are used. The spatial domain embedding features and the frequency domain embedding features are fused to obtain the embedding features of the low-frequency branch; S23. For high-frequency branches, the high-frequency components are embedded through convolution projection operations, and the embedded features are input into the attention mechanism, divided into query, key, value and additional features, and the attention weights are calculated. The weighted features are calculated and finally the weighted features are input into the feedforward neural network to obtain the embedded features of the high-frequency branches.

4. The method for visual detection of dental fluorosis according to claim 3, characterized in that: In the S3 step, the construction method of the low-frequency and high-frequency feature interaction module is as follows: the low-frequency embedded features and the high-frequency embedded features output by the preliminary enhancement module are input, the high-frequency embedded features are first segmented to obtain the horizontal component, vertical component and diagonal component of the high frequency, and then the coefficients of the three components are added to obtain the aggregated high-frequency embedded features, and the low-frequency embedded features and the aggregated high-frequency embedded features are mapped to the attention space using linear projection to obtain the query, key, low-frequency value and high-frequency value, calculate the attention weight, and perform weighted summation to obtain the low-frequency features and high-frequency features.

5. The method for enhancing dental fluorosis visual detection images according to claim 4, characterized in that: In the step S4, the construction method of the deepening enhancement module is: S41. Input the low-frequency features and high-frequency features output by the low- and high-frequency feature interaction module to obtain enhanced low-frequency components and high-frequency components. Calculate the residual between the real image and the enhanced low-frequency components. Calculate the residual between the real image and the enhanced high-frequency components. Use the two residuals as targets of the generator. Input the enhanced low-frequency components and high-frequency components to the generator. Combine the noise vector to generate restored frequency components. Use the inverse discrete wavelet transform to fuse the low-frequency components and high-frequency components generated by the generator to obtain an enhanced dental fluorosis image. S42. The loss function of the discriminator introduces perceptual loss and feature consistency loss on the basis of adversarial loss, and the overall loss of the discriminator is: ； Where, It's about fighting losses. is the feature consistency loss, and is a weight coefficient that controls the importance of perceptual loss and feature consistency loss in the total loss.

Images