Remote eye vision diagnosis and treatment intelligent service method and system and storage medium

Abstract

The application relates to the field of image processing, and discloses a remote eye optometry and treatment intelligent service method and system and a storage medium. The method comprises the following steps: acquiring an eye fundus image to be processed, and performing standardization processing on the image, wherein the standardization processing comprises normalizing a gray value of the image and performing geometric correction; performing edge detection processing on the image after the standardization, extracting a structural feature region contained in the image, and dividing an image subregion based on the structural feature region; the system comprises the following modules: an image acquisition and processing module, an encoding and transmission module, a decoding and recovery module, and an evaluation and diagnosis module. By introducing a region weight map and a multi-scale decomposition technology, the image reconstruction precision is significantly improved, the detail retention capability is optimized, higher-quality image restoration is realized, and the system performs excellently in complex images and scenes with rich details.

Description

Technical Field

[0001] This invention relates to the field of image processing, specifically to a method, system, and storage medium for remote optometry diagnosis and treatment intelligent services. Background Technology

[0002] In modern society, image data processing and analysis play a crucial role in various fields. For example, in applications such as medical imaging, video surveillance, and autonomous driving, accurately and quickly recovering image information while preserving image details and structure is a core issue. Image reconstruction technology is particularly vital in scenarios involving image compression, denoising, and missing data imputation. While existing image encoding and decoding technologies can accomplish basic reconstruction tasks, as demands continue to increase, handling the details of complex images and maintaining the integrity of important information remains an unsolved problem.

[0003] In existing technologies, variational autoencoders (VAEs), as a widely used image coding technique, have certain advantages in image compression and generation. By mapping an image into a latent space, VAEs can effectively compress image information while preserving, to some extent, the overall structural features of the image. Its advantage lies in its ability to perform efficient image reconstruction using a lower-dimensional latent representation, reducing computational resource consumption. In the decoding stage, VAEs employ standard upsampling and deconvolution techniques to reconstruct the approximate structure of the image.

[0004] However, existing technologies still have some shortcomings that limit their application in more complex image processing tasks. First, traditional VAE models often ignore the different importance of various regions in an image during feature encoding, leading to the easy loss or over-compression of information in key areas. Second, although some models attempt to introduce multi-scale processing, these methods still fail to effectively address the problem of representing image details and structure. In the decoding stage, existing technologies often rely on a single-scale reconstruction path, which makes it easy to distort local details and high-frequency information of the image, resulting in poor reconstructed image quality, especially in images with rich or diverse details. Furthermore, the training optimization of existing technologies usually focuses on pixel-level reconstruction loss, ignoring adaptive optimization at the structural level of the image, leading to the loss of structural information and blurring of details during image reconstruction. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a remote eye care intelligent service method, system, and storage medium, which solves the problems of neglecting regional information and insufficient multi-scale feature expression in the image reconstruction process of existing technologies.

[0006] To achieve the above objectives, the present invention provides a remote intelligent ophthalmology diagnosis and treatment service method, comprising the following steps:

[0007] Acquire a fundus image to be processed and perform standardization processing on the image. The standardization processing includes normalizing the gray values ​​of the image and geometric correction.

[0008] Edge detection processing is performed on the standardized image to extract regions containing structural features, and image sub-regions are divided based on the structural feature regions;

[0009] Information entropy values ​​are calculated for each of the image sub-regions, and a region weight map is generated based on the information entropy values, which is used for weighted configuration of compression processing of different image sub-regions;

[0010] The standardized image is feature-encoded based on a variational autoencoder. The region weight map is used to adjust the encoder’s latent spatial representation generation method for each image sub-region, and the corresponding latent coding data is generated.

[0011] The latent encoded data is used as input for decoding. Image features at different scales are extracted through multi-scale decomposition. The image features at each scale are then combined to perform variational decoding and reconstruction, generating the image reconstruction result.

[0012] Preferably, the normalization process and geometric correction specifically include:

[0013] The grayscale values ​​of the fundus image are normalized so that the pixel value range of the image is mapped to a specified grayscale range.

[0014] Geometric correction involves using affine and perspective transformations to correct distortions in an image caused by shooting angle and deformation.

[0015] Preferably, the structural feature region specifically includes:

[0016] The optic disc, macula, and vascular regions in fundus images contain high levels of texture information;

[0017] The extraction of the structural feature regions employs an edge detection algorithm to accurately extract key edges and contour information from the image.

[0018] Preferably, the information entropy value specifically includes:

[0019] The entropy value of each image sub-region is calculated using the following formula:

[0020] ;

[0021] in, The entropy value of the image region; Pixel values ​​within the image region The probability distribution; The quantity of pixel values; For the first image region Each pixel value;

[0022] Regions with high entropy values ​​contain more information;

[0023] During the calculation of the information entropy value, the probability distribution of each region is calculated using a grayscale histogram.

[0024] Preferably, the region weight map specifically includes:

[0025] A region weight map is generated based on the entropy value of each image sub-region.

[0026] The region weight map is generated in the following way:

[0027] ;

[0028] in, For position The region weight value at that location; For image region The entropy value; The maximum entropy value of all regions in the image;

[0029] Using the region weight map, different compression ratios are applied to different regions in subsequent compression steps.

[0030] Preferably, the feature encoding specifically includes:

[0031] The normalized image is encoded based on the variational autoencoder. The encoding process includes extracting deep features of the image through a convolutional neural network and generating a latent spatial representation through a fully connected layer.

[0032] During the feature encoding process, a variational inference method based on reparameterization techniques is used to optimize the distribution of the latent space.

[0033] Preferably, the potential encoded data specifically includes:

[0034] The latent spatial representation data generated in the encoding step represents low-dimensional features of the fundus image, and the feature encoding data consists of multiple encoding vectors;

[0035] The potential encoded data is further quantized to transform it into discrete data of a fixed size;

[0036] The latent encoded data is obtained by employing the latent space of a variational autoencoder. The probability distribution of the input image is represented by two parameters: mean and variance.

[0037] Preferably, the multi-scale decomposition method includes:

[0038] The input image is subjected to multi-scale wavelet transform, which decomposes the image into multiple frequency components. These frequency components are then encoded, decoded, and reconstructed.

[0039] The multi-scale decomposition method extracts information from the image at different scales;

[0040] The multi-scale decomposition employs discrete wavelet transform, which processes the low-frequency and high-frequency components of the image separately.

[0041] It also provides a remote eye care intelligent service system, including:

[0042] The image acquisition and processing module is used to acquire fundus images and perform image standardization processing, identify key regions, and perform weighted processing on the regions based on entropy values;

[0043] The encoding and transmission module encodes the weighted fundus image based on a variational autoencoder and transmits the compressed data to a remote terminal.

[0044] The decoding and recovery module is used to decode the received compressed data at the remote end and use multi-scale variational reconstruction technology to recover image details;

[0045] The assessment and diagnosis module is used to evaluate the quality of the restored images and provide image data to remote physicians to support ophthalmological diagnoses.

[0046] It also provides a storage medium storing a computer program, which, when executed by a processor, enables a remote intelligent service method for eye examination and treatment.

[0047] This invention provides a remote intelligent service method, system, and storage medium for optometry diagnosis and treatment. It has the following beneficial effects:

[0048] 1. This invention employs an image latent representation generation method based on variational autoencoders and introduces region weight maps for encoder guidance, achieving differentiated compressed representation of image structural regions. Compared to traditional non-discriminatory encoding schemes, it effectively avoids the loss of key structural information during the encoding process and solves the problem of uneven representation of local image details.

[0049] 2. This invention introduces a multi-scale decomposition method to extract image features during the decoding process, constructing a richer hierarchical information representation during structure restoration. Compared to existing techniques that only use a single-scale decoding path, this method improves the ability to restore textures and contours at different levels in complex images, avoiding problems such as overall image blurring or detail collapse.

[0050] 3. The joint strategy of region guidance and multi-scale reconstruction proposed in this invention enables images to have structure awareness in both the encoding and decoding stages. Compared with the traditional end-to-end black box feature compression, it improves the structural constraint mechanism in the latent space, effectively enhances the model's adaptability to semantic differences in image regions, and solves the technical shortcomings of information redundancy and compression distortion from the source.

[0051] 4. This invention introduces a scale-aware reconstruction loss function in the decoding stage and improves the model's ability to fit the original image structure through hierarchical alignment in the training stage. This approach differs from conventional image reconstruction mechanisms that rely solely on pixel-level MSE errors, making the model pay more attention to the consistency of the global structure and effectively overcoming the pain point of difficulty in recovering high-frequency information. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0053] Figure 2 This is a schematic diagram of the system architecture of the present invention. Detailed Implementation

[0054] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Please see the appendix Figure 1 This invention provides a remote eye care intelligent service method, including the following steps:

[0056] S1. Obtain the fundus image to be processed and perform standardization processing on the image. Standardization processing includes normalizing the gray values ​​of the image and geometric correction.

[0057] In remote optometry and intelligent service systems, image processing is one of the key preprocessing steps, especially the standardization process performed after acquiring the fundus images to be processed. The purpose of this step is to ensure that images from different sources have the same grayscale standard and geometric consistency, so as to provide reliable input for subsequent feature extraction, encoding, compression, and diagnosis.

[0058] Image standardization involves two key steps: normalizing the image's grayscale values ​​and performing geometric correction. These steps not only help eliminate inconsistencies caused by differences in equipment and environment during image acquisition but also ensure the geometric accuracy of the image, enabling precise analysis and diagnosis of fundus structures.

[0059] In this embodiment, after acquiring the fundus image to be processed, the image is first standardized, specifically including grayscale value normalization and geometric correction. This standardization process helps eliminate grayscale inconsistencies caused by external factors such as equipment and lighting, ensuring that the image has a uniform grayscale range in subsequent processing. Simultaneously, geometric correction is used to correct image shooting angle distortion and geometric distortion caused by the equipment, ensuring that the fundus structures in the image are accurately represented.

[0060] In this embodiment, the image grayscale values ​​are first normalized. Since different images may have different grayscale value ranges, normalization can map the image grayscale values ​​to a standard grayscale range to eliminate grayscale differences between images.

[0061] Generally, the grayscale range of fundus images varies depending on factors such as the acquisition device, shooting angle, and ambient lighting. To eliminate these differences, this embodiment employs a linear normalization method to adjust the image's grayscale values ​​to a uniform range. Specifically, let the original image grayscale value be... Its minimum and maximum values ​​are respectively and The grayscale values ​​of an image are transformed through linear transformation. Convert to normalized value :

[0062] ;

[0063] in, This refers to the normalized grayscale value. This method unifies images from different sources to the same grayscale range, thus eliminating inconsistencies in grayscale levels caused by differences in equipment or shooting conditions.

[0064] In some embodiments, the normalization process may further employ logarithmic transformation or histogram equalization to enhance the contrast of the image, thereby better highlighting the details of the fundus, especially in low-contrast areas.

[0065] Geometric correction of the image is another important standardization process. In acquired fundus images, the geometric structure may deviate due to differences in the shooting angle of the device or distortion caused by the image acquisition equipment. To ensure geometric consistency of the image, this embodiment employs geometric correction techniques.

[0066] Geometric correction can be achieved through various transformations, with common methods including affine transformations and perspective transformations. In one possible implementation, affine transformations are first used to correct distortions such as translation, rotation, and scaling in the image. The transformation matrix of an affine transformation... It can be represented as:

[0067] ;

[0068] in, , and , These are the coordinates of a point in the original image and the corrected image, respectively. , , , These are the parameters in the affine transformation matrix; , This is the translation amount. This affine transformation can handle rotation and translation in images and ensure spatial alignment of the images.

[0069] For more complex geometric distortions, this embodiment also employs perspective transformation. Perspective transformation is used to handle image perspective distortion caused by incorrect shooting angles or the optical characteristics of the device. The transformation matrix for perspective transformation is... as follows:

[0070] ;

[0071] in, , and , These are the coordinates of a point in the original image and the corrected image, respectively. - The coefficients of each element in the perspective transformation matrix represent the combination of transformation factors such as scaling, rotation, translation, and projection.

[0072] Perspective transformation can correct geometric distortions in an image, restore the actual perspective of the image, and make the fundus structure in the image more realistic.

[0073] As an alternative, feature-matching-based geometric correction methods can also be employed. Image alignment is achieved by extracting feature points from the image, such as corner points and edge points, and using these feature points for registration. Commonly used feature extraction algorithms include SIFT (Scale Invariant Feature Transform) and SURF (Speed-Up Robust Feature Transform). After feature matching, the least squares optimization algorithm is used to calculate the image's transformation matrix, which is then applied to the image's geometric correction.

[0074] S2. Perform edge detection processing on the standardized image, extract regions containing structural features, and divide the image into sub-regions based on the structural feature regions;

[0075] After standardizing the fundus images, edge detection is necessary to further locate key visual structural regions within the images. This process aims to extract regions containing structural features, such as the optic disc, blood vessels, and macula, and to divide the image into several sub-regions based on these features for subsequent feature extraction, image encoding, or lesion identification.

[0076] In general, accurate identification of structural edges can help determine local areas in an image that are of medical diagnostic significance. In actual processing, edge detection not only undertakes the task of locating structural regions, but also provides basic spatial information to support the division of image sub-regions.

[0077] In this embodiment, immediately after image normalization is completed, edge detection is performed on the normalized image to obtain structural feature boundaries in the image. This edge detection can be achieved based on the gradient intensity of image grayscale changes. By calculating the grayscale changes of each pixel in the image in the horizontal and vertical directions, locations with abrupt grayscale changes are identified as potential edge points.

[0078] Specifically, let the standardized image be... ,in Given the two-dimensional spatial coordinates of the image, edge detection can be achieved through gradient calculation as follows:

[0079] ;

[0080] ;

[0081] in, Indicates the image in The gradient value in the direction; This represents the gradient value in the direction; This is the gradient magnitude at that point, used to measure the edge strength at that location; For the image in Partial derivatives in the direction; For the image in Partial derivatives in the direction.

[0082] In one possible implementation, edge detection can be performed using the Sobel operator, the Prewitt operator, or the Canny edge detection method. As an alternative, Canny edge detection is preferred in this embodiment due to its high edge localization accuracy and noise resistance. The Canny algorithm comprises four main steps: Gaussian filtering, gradient calculation, non-maximum suppression, and double-threshold edge connection.

[0083] After edge detection, the extracted edge regions are further processed to identify image regions with structural features. For example, based on closed edge contours, spatial regions containing the optic disc boundary, the main blood vessel orientation, and the macular dark area are extracted. To ensure the accuracy of structural regions, region growing algorithms or contour tracking methods based on edge morphology features can be used to determine edge connectivity.

[0084] In some embodiments, to enhance the structure awareness capability of sub-region partitioning, an edge-guided watershed algorithm can be introduced, combined with gradient maps. Construct a watershed marker map to achieve stable segmentation of structural boundaries.

[0085] As an extension, in the process of dividing an image into sub-regions, statistical characteristics such as pixel density, average gray level, and edge complexity within the region can be combined for optimization and judgment to ensure the spatial closure and anatomical rationality of the divided sub-regions.

[0086] Specifically, after the sub-regions are divided, each sub-region can be labeled and numbered for use in subsequent region selection, local feature extraction, or image compression processing. Each sub-region can also be accompanied by its edge morphology information, such as perimeter and boundary curvature, for structural complexity analysis.

[0087] S3. Calculate the information entropy value for each sub-region of the image, and generate a region weight map based on the information entropy value, which is used for weighted configuration of compression processing of different sub-regions of the image.

[0088] After extracting and subdividing the image's structural features, a more targeted image compression strategy requires evaluating the differences in information representation value among the sub-regions. This invention introduces an information entropy measurement mechanism to quantitatively reflect the complexity and detail richness of image content in different sub-regions, and constructs a region weight map based on the information entropy values. This region weight map is then used to control the weighting configuration during image compression, thereby optimizing the strategy by retaining more image information in important regions and applying higher compression ratios to less important regions.

[0089] Generally, regions in an image containing rich texture or structural details have a more dispersed pixel grayscale distribution and are relatively more important to the overall image representation. These regions often have higher information entropy values. Therefore, region weighting based on information entropy values ​​is an effective compression weighting method driven by image statistical characteristics.

[0090] In this embodiment, after the image sub-regions are divided, the information entropy value corresponding to each image sub-region is calculated.

[0091] Regional weight maps can be generated in the following ways:

[0092] The entropy value of each image sub-region is calculated using the following formula:

[0093] ;

[0094] in, The entropy value of the image region; Pixel values ​​within the image region The probability distribution; The quantity of pixel values; For the first image region Each pixel takes a value.

[0095] Alternatively, if the image is a color image, the information entropy can be calculated for each channel separately, and then a weighted average can be used to obtain the overall information entropy:

[0096] ;

[0097] in, , , These represent the information entropy in the red, green, and blue channels, respectively. , , These are the weighting coefficients for each channel.

[0098] After calculating the information entropy of all sub-regions, a regional weight map is further generated. This weight map is used to guide the compression ratio configuration adopted for each region during the compression process. Specifically, let the weight map be... Its definition is as follows:

[0099] ;

[0100] in, For position The region weight value at that location; For image region The entropy value; This represents the maximum entropy value across all regions in the image.

[0101] As an implementation strategy, in the image compression module, for each sub-region, based on its corresponding weight value... The compression parameters are dynamically adjusted. For example, in compression algorithms based on transform coding (such as DCT or DWT), more high-frequency coefficients are retained to reduce the loss of structural regions, while higher quantization steps are applied to low-weight regions.

[0102] In some embodiments, the region weight map can be used to generate a corresponding compression parameter mapping table, which can be directly called by the subsequent image encoder. This mapping table contains the compression level identifier, quantization parameters, and precision retention level for each region, ensuring that the compression algorithm can be flexibly configured according to the importance of each region.

[0103] S4. Based on the variational autoencoder, feature encoding is performed on the standardized image. The region weight map is used to adjust the encoder's latent spatial representation generation method for each image sub-region, and the corresponding latent coding data is generated.

[0104] After extracting structural feature regions, dividing sub-regions, and constructing information entropy weight maps for standardized images, this invention further introduces an image compression coding mechanism based on a variational autoencoder (VAE) to achieve adaptive image feature coding. This mechanism generates a structure by learning the latent spatial representation of the image. While compressing and preserving key image feature information, it adjusts the encoder's region-awareness capability in conjunction with the aforementioned generated region weight map, thereby achieving differentiated feature abstraction and latent coding optimization.

[0105] Traditional image coding methods often fail to fully utilize the semantic differences between regions in an image, leading to the loss of information in structurally important regions or invalid and redundant coding. This invention introduces a variational autoencoder and combines it with a region weight map to control the encoder's perception and coding granularity for each sub-region, making feature extraction and compression coding more content-sensitive and adaptable.

[0106] In this embodiment, based on the completed image normalization processing, structural edge extraction, sub-region segmentation, and region weight map construction, the image is input into a variational autoencoder model for feature encoding. The variational autoencoder mainly consists of two parts: an encoder network and a decoder network. The encoder maps the input image to a set of distribution parameters in the latent space, which typically include a mean vector and a log-variance vector.

[0107] Specifically, noise vectors are sampled from a standard normal distribution using the reparameter resampling technique. And compute the latent representation :

[0108] ;

[0109] in, and These are the mean and standard deviation vectors of the encoder output, respectively. This represents the sampling noise term.

[0110] In one possible implementation, to make the latent representation generated by the encoder more structurally sensitive in terms of spatial distribution, this invention utilizes the aforementioned region weight map to regulate the encoder structure. During the image input to the encoder, the region weight map is introduced as an additional guiding input and fused with the image feature map. For example, a weight map is introduced in the feature extraction layer of the encoder. The feature map is weighted and superimposed channel-wise or pixel-wise to improve the model's ability to perceive high-weight regions.

[0111] As an alternative, during the encoder feature extraction stage, the weight map can be concatenated into the input image feature tensor with extended channels, allowing the model to automatically learn the influence of region weighting mechanisms on the latent representation distribution during the encoding process. This approach can enhance the spatial selectivity of the model while maintaining its structural stability.

[0112] In this embodiment, the latent representation output by the encoder Designed as multi-dimensional feature vectors, they can compress and represent important structural information of the entire image. For different sub-regions in the image, the region weights in the latent space are reflected as enhanced distribution constraints on the dimensions of specific regions. That is, the latent encoding dimension distribution of high-weight regions tends to be concentrated and the information density is higher, thus exhibiting better reconstruction ability in subsequent reconstruction.

[0113] In some embodiments, to improve the model's accuracy in encoding latent features in local regions, a region attention mechanism can be set in the intermediate layer of the encoder to automatically enhance the network's expressive ability in key regions based on the image weight map. This attention mechanism makes the model more sensitive to feature learning in high-information-entropy regions by dynamically adjusting the response values ​​of the intermediate layer feature channels.

[0114] Furthermore, during the latent space distribution fitting process, this invention maintains the core optimization objective of VAE, namely maximizing the variational lower bound (ELBO, Evidence Lower Bound). The optimization objective includes a reconstruction error term and a KL divergence term, which respectively measure the difference between the reconstructed image and the input image, and the difference between the latent distribution and the standard normal distribution. The introduction of a weight map can further adjust the region-aware function in this optimization objective, thereby achieving backpropagation gradient adjustment of the reconstruction error based on region importance.

[0115] S5. The latent encoded data is used as input for decoding. Image features at different scales are extracted through multi-scale decomposition. The image features at each scale are combined to perform variational decoding and reconstruction to generate the image reconstruction result.

[0116] After completing the feature encoding of the standardized image and generating a latent spatial representation based on the region weight map, decoding and reconstruction processing of the latent encoded data is required to further recover image information and meet the input requirements of subsequent visual analysis or diagnostic applications. This invention introduces a multi-scale decomposition and fusion mechanism, which extracts image features at multiple scales during the decoding stage and performs hierarchical reconstruction by combining features from each scale to enhance the structural expressiveness and detail preservation of the reconstructed image. This processing strategy not only achieves efficient integration with the aforementioned encoding module but also provides a structure-aware foundation for quality control of the reconstructed image.

[0117] Generally, images exhibit different types of information at different scales. Low-scale levels mainly contain global structural features, while high-scale levels highlight edges, textures, and local details. Therefore, introducing multi-scale processing mechanisms during image decoding helps to more comprehensively recover the information content carried in the latent representation in both the spatial and frequency domains.

[0118] In this embodiment, the latent encoded data generated in the previous stage is used as the decoding input, and a reparameter sampling operation is performed on the latent representation to obtain the latent vector for decoding. The latent vector is first subjected to multiple upsampling operations (such as deconvolution or pixel rearrangement modules) in the decoder network for preliminary image reconstruction. To improve decoding quality, a multi-scale decomposition module is introduced into the decoder structure to extract intermediate features layer by layer during the decoding process, forming image features at multiple scale levels.

[0119] As an alternative, based on the scale extraction described above, the image features at each scale can be cascaded, weighted, or fused using a feature fusion module. This process can utilize channel attention or spatial attention mechanisms to achieve adaptive weight learning across scales, thereby enhancing the reconstructive representation of key regions.

[0120] Specifically, in the fusion operation, feature maps at each scale can be set. Corresponding to a learnable fusion weight Final image feature representation It can be represented as:

[0121] ;

[0122] in, This represents the total number of feature layers involved in loss calculation within the network.

[0123] In some embodiments, in order to maintain structural coherence, an upsampling interpolation operation can be introduced before the fused features are fed into the last layer decoding module to keep the feature map size consistent across all scales, which facilitates subsequent convolution to restore the final image size.

[0124] In this embodiment, the fused image features are further processed by the decoding tail module to finally generate the image reconstruction result. This reconstruction result has the same size in space as the standardized image and is optimized through reconstruction error control mechanisms (such as MSE loss, SSIM loss, or perceptual loss) during the training process to achieve the expected balance between structural and semantic reconstruction.

[0125] As an implementation strategy, the loss function design in the variational decoder structure can introduce a scale-aware error term on top of the reconstruction error. This term measures whether image features at each scale are effectively recovered during the reconstruction process. It can be accomplished by performing the same scale decomposition on both the original image and the reconstructed image, and calculating the error values ​​between corresponding levels.

[0126] For example, let the original graph be... Layer features are The features of the corresponding layer in the reconstructed image are Then the scale perception loss is:

[0127] ;

[0128] in, The scale-aware loss function; This represents the total number of feature layers in the network that participate in loss calculation. For the first The weighting coefficients of the layers.

[0129] This loss term, together with the traditional image reconstruction loss function, constitutes the final optimization objective function, ensuring the effective alignment and restoration of image features at different scales during the decoding process.

[0130] The remote optometry diagnosis and treatment intelligent service system described below can be referred to in correspondence with the remote optometry diagnosis and treatment intelligent service method described above.

[0131] Please see the appendix Figure 2 The present invention also provides a remote optometry diagnosis and treatment intelligent service system, including:

[0132] The image acquisition and processing module is used to acquire fundus images and perform image standardization processing, identify key regions, and perform weighted processing on the regions based on entropy values;

[0133] The encoding and transmission module encodes the weighted fundus image based on a variational autoencoder and transmits the compressed data to a remote terminal.

[0134] The decoding and recovery module is used to decode the received compressed data at the remote end and use multi-scale variational reconstruction technology to recover image details;

[0135] The assessment and diagnosis module is used to evaluate the quality of the restored images and provide image data to remote physicians to support ophthalmological diagnoses.

[0136] The system in this embodiment can be used to execute the above method embodiments, and its principle and technical effect are similar, so they will not be described again here.

[0137] The present invention also provides a storage medium storing a computer program, which is executed by a processor to perform the method described above.

[0138] The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0139] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A remote intelligent service method for optometry diagnosis and treatment, characterized in that, Includes the following steps: Acquire a fundus image to be processed and perform standardization processing on the image. The standardization processing includes normalizing the gray values ​​of the image and geometric correction. Edge detection processing is performed on the standardized image to extract regions containing structural features, and image sub-regions are divided based on the structural feature regions; Information entropy values ​​are calculated for each of the image sub-regions, and a region weight map is generated based on the information entropy values, which is used for weighted configuration of compression processing of different image sub-regions; The standardized image is feature-encoded based on a variational autoencoder. The region weight map is used to adjust the encoder’s latent spatial representation generation method for each image sub-region, and the corresponding latent coding data is generated. The latent encoded data is used as input for decoding. Image features at different scales are extracted through multi-scale decomposition. The image features at each scale are combined to perform variational decoding and reconstruction to generate the image reconstruction result. The information entropy value specifically includes: The entropy value of each image sub-region is calculated using the following formula: ； in, The entropy value of the image region; Pixel values ​​within the image region The probability distribution; The number of pixel values; For the first image region Each pixel value; Regions with high entropy values ​​contain more information; During the calculation of the information entropy value, the probability distribution of each region is calculated using a grayscale histogram. The region weight map specifically includes: A region weight map is generated based on the entropy value of each image sub-region. The region weight map is generated in the following way: ； in, For position The region weight value at that location; For image region The entropy value; The maximum entropy value of all regions in the image; Using the region weight map, different compression ratios are applied to different regions in subsequent compression steps.

2. The remote optometry diagnosis and treatment intelligent service method according to claim 1, characterized in that, The normalization process and geometric correction specifically include: The grayscale values ​​of the fundus image are normalized so that the pixel value range of the image is mapped to a specified grayscale range. Geometric correction involves using affine and perspective transformations to correct distortions in an image caused by shooting angle and deformation.

3. The remote optometry diagnosis and treatment intelligent service method according to claim 1, characterized in that, The structural feature regions specifically include: The optic disc, macula, and vascular regions in fundus images contain high levels of texture information; The extraction of the structural feature regions employs an edge detection algorithm to accurately extract key edges and contour information from the image.

4. The remote optometry diagnosis and treatment intelligent service method according to claim 1, characterized in that, The feature encoding specifically includes: The normalized image is encoded based on the variational autoencoder. The encoding process includes extracting deep features of the image through a convolutional neural network and generating a latent spatial representation through a fully connected layer. During the feature encoding process, a variational inference method based on reparameterization techniques is used to optimize the distribution of the latent space.

5. The remote optometry diagnosis and treatment intelligent service method according to claim 1, characterized in that, The potential encoded data specifically includes: The latent spatial representation data generated in the encoding step represents low-dimensional features of the fundus image, and the feature encoding data consists of multiple encoding vectors; The potential encoded data is further quantized to transform it into discrete data of a fixed size; The latent encoded data is obtained by employing the latent space of a variational autoencoder. The probability distribution of the input image is represented by two parameters: mean and variance.

6. The remote optometry diagnosis and treatment intelligent service method according to claim 1, characterized in that, The multi-scale decomposition method includes: The input image is subjected to multi-scale wavelet transform, which decomposes the image into multiple frequency components. These frequency components are then encoded, decoded, and reconstructed. The multi-scale decomposition method extracts information from the image at different scales; The multi-scale decomposition employs discrete wavelet transform, which processes the low-frequency and high-frequency components of the image separately.

7. A remote optometry diagnosis and treatment intelligent service system, comprising the remote optometry diagnosis and treatment intelligent service method according to any one of claims 1-6, characterized in that, include: The image acquisition and processing module is used to acquire fundus images and perform image standardization processing, identify key regions, and perform weighted processing on the regions based on entropy values; The encoding and transmission module encodes the weighted fundus image based on a variational autoencoder and transmits the compressed data to a remote terminal. The decoding and recovery module is used to decode the received compressed data at the remote end and use multi-scale variational reconstruction technology to recover image details; The assessment and diagnosis module is used to evaluate the quality of the restored images and provide image data to remote physicians to support ophthalmological diagnoses.

8. A storage medium, characterized in that, The device contains a computer program that, when executed by a computer, implements the remote optometry diagnosis and treatment intelligent service method according to any one of claims 1-6.

Images