Artificial intelligence-based tongue image extraction method and device, computer equipment and medium

Abstract

The present application relates to the technical field of artificial intelligence, and especially relates to a tongue image extraction method and device based on artificial intelligence, computer equipment and medium. The method extracts initial image features of a tongue image by using a feature extraction layer, convolves the initial image features by using a first convolution kernel and a second convolution kernel in turn, obtains high-dimensional information features, divides the high-dimensional information features into feature subblocks, determines the relative positions of each feature subblock, forms a subblock sequence by using feature subblocks representing the same relative position, encodes the subblock sequence into an encoding sequence, determines the encoding results of the feature subblocks from the encoding sequence, restores all the encoding results into encoding features, pools the encoding features into pooled features, decodes the pooled features, determines the decoding results as tongue image segmentation results, and makes the feature subblocks contain local information by convolution, encodes the subblock sequence composed of the feature subblocks, and makes the encoding features contain local information and global information at the same time, thereby improving the tongue image segmentation effect.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, computer equipment, and medium for tongue image extraction based on artificial intelligence. Background Technology

[0002] With the development of artificial intelligence (AI) technology, deep learning models have been widely used in various application tasks in the field of smart healthcare, such as medical image segmentation, medical image classification, and medical image detection. In medical image segmentation, taking tongue image segmentation as an example, deep learning models based on convolutional neural networks are usually used to achieve semantic segmentation of tongue images.

[0003] However, deep learning models based on convolutional neural networks (CNNs) suffer from limitations in feature extraction due to the convolution kernel size, resulting in relatively limited features that fail to effectively represent the global feature information of tongue images. Furthermore, CNN-based deep learning models struggle to capture the relationships between features in distant regions within a tongue image, leading to noise points and even segmentation errors in the segmentation results, resulting in poor accuracy. Therefore, improving the accuracy of tongue image segmentation has become an urgent problem to be solved. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a method, apparatus, computer device and medium for tongue image extraction based on artificial intelligence, in order to solve the problem of poor accuracy in tongue image segmentation.

[0005] In a first aspect, embodiments of the present invention provide a tongue image extraction method based on artificial intelligence, the tongue image extraction method comprising:

[0006] The tongue image to be segmented is input into the feature extraction layer for feature extraction to obtain the initial image features;

[0007] The initial image features are convolved using a first convolution kernel to obtain local information features. The local information features are then convolved using a second convolution kernel to obtain high-dimensional information features. The high-dimensional information features and the local information features have the same feature size, and the number of channels in the high-dimensional information features is greater than the number of channels in the local information features.

[0008] The high-dimensional information feature is divided into N non-overlapping feature blocks. For any feature block, the feature block is divided into M non-overlapping feature sub-blocks. The relative position of each feature sub-block in the feature block is determined. The number of channels of the feature block and the number of channels of the feature sub-block are the same as the number of channels of the high-dimensional information feature. N and M are both positive integers.

[0009] According to the preset query order, feature sub-blocks representing the same relative position in all feature blocks are formed into a sub-block sequence, and the sub-block sequence corresponding to the relative position is obtained. The encoder is used to encode each sub-block sequence to obtain the encoding sequence of the corresponding sub-block sequence. The encoding result of each feature sub-block is determined from the corresponding encoding sequence.

[0010] Based on the relative positions of all feature sub-blocks, all encoding results are restored to the encoding features corresponding to the feature blocks. Pooling is performed on the encoding features to obtain pooled features. Feature decoding is performed based on the pooled features to obtain the decoding result. The decoding result is determined to be the tongue image segmentation result.

[0011] Secondly, embodiments of the present invention provide an artificial intelligence-based tongue image extraction device, the tongue image extraction device comprising:

[0012] The feature extraction module is used to input the tongue image to be segmented into the feature extraction layer for feature extraction to obtain the initial image features;

[0013] The convolution processing module is used to perform convolution processing on the initial image features using a first convolution kernel to obtain local information features, and to perform convolution processing on the local information features using a second convolution kernel to obtain high-dimensional information features. The high-dimensional information features and the local information features have the same feature size, and the number of channels of the high-dimensional information features is greater than the number of channels of the local information features.

[0014] The feature segmentation module is used to divide the high-dimensional information feature into N non-overlapping feature blocks. For any feature block, the feature block is divided into M non-overlapping feature sub-blocks. The relative position of each feature sub-block in the feature block is determined. The number of channels of the feature block and the number of channels of the feature sub-block are the same as the number of channels of the high-dimensional information feature. N and M are both positive integers.

[0015] The sequence encoding module is used to form a sub-block sequence by forming the feature sub-blocks that represent the same relative position in all feature blocks according to a preset query order, thereby obtaining the sub-block sequence of the corresponding relative position. The encoder is used to encode each sub-block sequence to obtain the encoding sequence of the corresponding sub-block sequence, and the encoding result of each feature sub-block is determined from the corresponding encoding sequence.

[0016] The tongue image segmentation module is used to restore all encoding results to the encoding features corresponding to the feature blocks according to the relative positions of all feature sub-blocks, perform pooling operation on the encoding features using a pooling layer to obtain pooled features, perform feature decoding based on the pooled features to obtain decoding results, and determine the decoding results as tongue image segmentation results.

[0017] Thirdly, embodiments of the present invention provide a computer device, the computer device including a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the tongue image extraction method as described in the first aspect.

[0018] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the tongue image extraction method as described in the first aspect.

[0019] The beneficial effects of the embodiments of the present invention compared with the prior art are as follows:

[0020] The tongue image to be segmented is input into the feature extraction layer for feature extraction to obtain initial image features. The initial image features are then convolved using a first convolution kernel to obtain local information features. A second convolution kernel is then used to convolve the local information features to obtain high-dimensional information features. These high-dimensional information features are divided into N non-overlapping feature blocks. For any feature block, it is further divided into M non-overlapping feature sub-blocks. The relative position of each feature sub-block within the feature block is determined. Following a preset query order, feature sub-blocks representing the same relative position from all feature blocks are grouped into a sub-block sequence, resulting in a sub-block sequence with corresponding relative positions. An encoder is then used to encode each sub-block sequence, yielding the encoded sequence for the corresponding sub-block sequence. The encoding result of each feature sub-block is determined from the corresponding encoding sequence. According to the relative position of all feature sub-blocks, all encoding results are restored to the encoding features of the corresponding feature blocks. Pooling is performed on the encoding features to obtain pooled features. Feature decoding is performed based on the pooled features to obtain the decoding result. The decoding result is determined to be the tongue image segmentation result. Local information features are extracted through convolution processing, and then the features are segmented so that each feature sub-block contains the local information of its respective feature block. Then, the encoder encodes the sub-block sequence at the same position to achieve the extraction of global information. This makes the encoded features contain both local and global information, improving the richness and representation ability of the encoded features, and thus improving the effect of tongue image segmentation. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1This is a schematic diagram of an application environment for an artificial intelligence-based tongue image extraction method provided in Embodiment 1 of the present invention;

[0023] Figure 2 This is a flowchart illustrating an artificial intelligence-based tongue image extraction method provided in Embodiment 1 of the present invention.

[0024] Figure 3 This is a schematic diagram of the structure of a tongue image extraction device based on artificial intelligence provided in Embodiment 2 of the present invention;

[0025] Figure 4 This is a schematic diagram of the structure of a computer device provided in Embodiment 3 of the present invention. Detailed Implementation

[0026] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0027] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0028] It should also be understood that the term “and / or” as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0029] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0030] Furthermore, in the description of this invention and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0031] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of the invention include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0032] The embodiments of this invention can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0033] Foundational technologies for artificial intelligence generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0034] It should be understood that the sequence number of each step in the following embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0035] To illustrate the technical solution of the present invention, specific embodiments are described below.

[0036] The tongue image extraction method based on artificial intelligence provided in Embodiment 1 of this invention can be applied to, for example... Figure 1In this application environment, the client communicates with the server. The client includes, but is not limited to, handheld computers, desktop computers, laptops, ultra-mobile personal computers (UMPCs), netbooks, cloud terminal devices, and personal digital assistants (PDAs). The server can be a standalone server or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0037] Both the client and server can be deployed in medical scenarios, such as hospitals and medical institutions. The client can acquire tongue images input by users, such as doctors and nurses. The client can also connect to image acquisition devices to directly acquire tongue images of target objects, such as patients seeking medical treatment.

[0038] See Figure 2 This is a flowchart illustrating an artificial intelligence-based tongue image extraction method provided in Embodiment 1 of the present invention. The above-described tongue image extraction method can be applied to... Figure 1 The server-side component connects to the client computer to acquire the tongue image to be segmented. The server-side computer contains a feature extraction layer and a pooling layer. The feature extraction layer extracts image features from the tongue image, while the pooling layer performs dimensionality reduction on the input features to aggregate feature information. For example... Figure 2 As shown, the tongue image extraction method may include the following steps:

[0039] Step S201: Input the tongue image to be segmented into the feature extraction layer for feature extraction to obtain the initial image features.

[0040] Among them, the tongue image can refer to image data containing the tongue, the feature extraction layer can be used to extract the image feature information of the tongue image, and the initial image features can be used to characterize the feature information of the tongue image.

[0041] Specifically, the feature extraction layer can be implemented using a convolutional neural network. In this embodiment, since the performance of a convolutional neural network degrades as its depth increases, and the depth can be represented by the number of convolutional layers, a residual neural network is used as the feature extraction layer. The residual neural network can map the feature information of the shallow layer to the deep network through skip connections, avoiding gradient vanishing or gradient exploding.

[0042] Optionally, the feature extraction layer includes a first local convolutional layer, a first channel convolutional layer, and a batch normalization layer;

[0043] The tongue image to be segmented is input into the feature extraction layer for feature extraction, resulting in initial image features including:

[0044] The tongue image is input into the first local convolutional layer and the first channel convolutional layer respectively, and both are processed by the batch normalization layer to obtain the first convolution result and the second convolution result;

[0045] The first convolution result and the second convolution result are added together, and the result is determined as the first convolution feature. The first convolution feature is used as the initial image feature.

[0046] The first local convolutional layer can be used to extract local feature information of the input information. In this embodiment, the first local convolutional layer can be processed by a 3*3 convolutional kernel. The first channel convolutional layer can be used to aggregate feature information of different channels and / or control the dimension of the output feature, that is, the number of channels. The first channel convolutional layer can be processed by a 1*1 convolutional kernel.

[0047] Batch normalization layers can be used to improve the training and convergence speed of networks, prevent gradient explosion or gradient vanishing, and avoid overfitting.

[0048] The first convolution result can represent the deep features corresponding to the tongue image, and the second convolution result can represent the mapping result of the shallow features corresponding to the tongue image. It should be noted that the deep features and shallow features here are only features that are deeper and shallower than each other. The first convolution feature can represent the result after the fusion of shallow features and deep features.

[0049] Specifically, preliminary feature extraction can be performed on the tongue image first. For example, a 7*7 convolutional kernel can be used to process the tongue image. Then, the processing results are sequentially pooled and convolved using a 1*1 convolutional kernel to obtain initial layer features. These initial layer features are then input into the first local convolutional layer and the first channel convolutional layer, and both are processed by a batch normalization layer to obtain the first convolution result and the second convolution result. The first convolution result and the second convolution result are added together, and the sum is determined as the first convolution feature. It should be noted that in this embodiment, the size of the first convolution feature is half of the initial layer feature, that is, the width of the first convolution feature is half the width of the initial layer feature, and the height of the first convolution feature is half the height of the initial layer feature.

[0050] After adding the results of the first and second convolutions, an activation function can be used to process the result. The result after activation function processing is then determined as the first convolution feature. Activation functions such as ReLU and Sigmoid can be used.

[0051] In one implementation, after obtaining the first convolutional feature corresponding to the initial layer features, a 1*1 size convolutional kernel can be used to convolve the first convolutional feature. The result of this convolutional processing is then fused with the result of the initial layer features after convolution with a 1*1 size convolutional kernel. The fused result is used as the initial image feature, thereby realizing nested short-distance skip connections and long-distance skip connections. Here, the distance can be represented by the difference in the number of convolutional layers, so that the shallow features are more fully and comprehensively mapped into the deep network, further improving the feature extraction capability of the network and the representation capability of the features extracted by the deep network.

[0052] In this embodiment, for cases where the width and height of the input features are twice that of the output features, the input features are processed through a first local convolutional layer and a first channel convolutional layer, respectively, and the processing results are directly added together to obtain the first convolutional feature. Compared to the conventional residual network method of adding the result of the input features processed by two local convolutional layers sequentially to the result of the input features processed by the channel convolutional layer to obtain the first convolutional feature, fewer convolutional operations are required, which speeds up the frequency of shallow information mapping to deep networks. Furthermore, a batch normalization layer is used after each convolutional processing, which is more effective than the conventional residual network method of only using a batch normalization layer on the result of the local convolutional layer processing. This more effectively avoids gradient vanishing or gradient exploding, thereby effectively avoiding network degradation and improving the performance of the feature extraction layer.

[0053] Optionally, the feature extraction layer may also include a second local convolutional layer and a second channel convolutional layer;

[0054] After determining that the summation result is the first convolutional feature, the following steps are also included:

[0055] The first convolutional features are input into the second local convolutional layer and the second channel convolutional layer, respectively, and both are processed by the batch normalization layer to obtain the third convolutional result and the fourth convolutional result.

[0056] The first convolutional feature is input into the batch normalization layer for processing to obtain the normalization result. The third convolutional result, the fourth convolutional result and the normalization result are added together to determine the added result as the second convolutional feature.

[0057] Accordingly, using the first convolutional features as initial image features includes:

[0058] The second convolutional feature is used as the initial image feature.

[0059] The second local convolutional layer can be used to extract local feature information of the input information. In this embodiment, the second local convolutional layer can also be processed by a 3*3 convolutional kernel. The second channel convolutional layer can be used to aggregate feature information of different channels and / or control the dimension of the output features. The second channel convolutional layer can be processed by a 1*1 convolutional kernel.

[0060] The third convolution result can represent the deep features corresponding to the tongue image, and the fourth convolution result can represent the mapping result of the shallow features corresponding to the tongue image. It should be noted that the deep features and shallow features here are only the deeper and shallower features relative to each other. The second convolution feature can represent the result after the fusion of the above shallow features and deep features.

[0061] Specifically, preliminary feature extraction can be performed on the first convolutional feature. For example, a 1*1 size convolutional kernel can be used to process the first convolutional feature. The processing results are then input into the second local convolutional layer and the second channel convolutional layer, and both are processed by a batch normalization layer to obtain the third and fourth convolutional results. The first convolutional feature is then input into the batch normalization layer for processing to obtain the normalized result. The third convolutional result, the fourth convolutional result, and the normalized result are then added together to determine the sum as the second convolutional feature. It should be noted that in this embodiment, the size of the second convolutional feature is the same as the size of the first convolutional feature.

[0062] After adding the results of the third convolution, the fourth convolution, and the normalization, an activation function can be used to process the summed result. The result after activation function processing is determined as the second convolution feature. The activation function can be ReLU, Sigmoid, or similar functions.

[0063] In one implementation, after obtaining the second convolution feature corresponding to the first convolution feature, a 1*1 size convolution kernel can be used to perform convolution processing on the second convolution feature. The result of the convolution processing is then fused with the first convolution feature, and the fused result is used as the initial image feature, thereby realizing nested short-distance skip-layer connections and long-distance skip-layer connections.

[0064] In one embodiment, the above-described structure of a second local convolutional layer, a second channel convolutional layer, and a skin normalization layer can be used for iterative processing. The input feature of the initial iteration is the first convolutional feature, and the output is the iteration result. The iteration result is used as the input feature for the next iteration until the iteration meets a preset value, at which point the iteration stops. The final output result is used as the second convolutional feature, which is also used as the initial image feature. The preset value can be set to 2. In this embodiment, through multi-layer iterative processing with unchanged input and output feature sizes, deeper feature information can be extracted, thereby improving the representational ability of the initial image features.

[0065] In this embodiment, for cases where the width and height of the input features are the same as the output features, the input features are processed through a second local convolutional layer and a second channel convolutional layer, respectively. The processing result is then directly added to the batch normalized result of the input features to obtain the second convolutional feature. Compared to conventional residual networks, which process the input features sequentially through two local convolutional layers and then add the result to the input features to obtain the second convolutional feature, fewer convolutional operations are required, accelerating the mapping of shallow information to deep networks. Furthermore, a batch normalization layer is used after each convolutional processing, which, compared to conventional residual networks that only use a batch normalization layer on the result of local convolutional layer processing, more effectively avoids gradient vanishing or gradient exploding, thereby effectively preventing network degradation and improving the performance of the feature extraction layer.

[0066] The above steps of inputting the tongue image to be segmented into the feature extraction layer to extract features and obtain initial image features provide initial image features that characterize the tongue image information for subsequent tongue image segmentation processing, which facilitates accurate tongue image segmentation.

[0067] Step S202: The initial image features are convolved using the first convolution kernel to obtain local information features, and the local information features are convolved using the second convolution kernel to obtain high-dimensional information features.

[0068] In this embodiment, the high-dimensional information features and the local information features have the same feature size, and the number of channels in the high-dimensional information features is greater than the number of channels in the local information features. The first convolutional kernel can be used to extract local information, and in this embodiment, the first convolutional kernel adopts a 3*3 size. The second convolutional kernel can be used to extract higher-dimensional local information, and in this embodiment, the second convolutional kernel adopts a 1*1 size.

[0069] Specifically, the initial image can be processed using the network structure where the input feature width and height are twice the output feature width and height. The processed result can be convolved using the first convolution kernel to obtain local information features.

[0070] The network structure may include two 1*1 convolutional layers, a first local convolutional layer, a first channel convolutional layer, and a batch normalization layer. The processing procedure for the network structure is described in the previous steps and will not be repeated here.

[0071] The steps described above, which involve using a first convolution kernel to convolve the initial image features to obtain local information features, and then using a second convolution kernel to convolve the local information features to obtain high-dimensional information features, represent local information with high-dimensional information features. This facilitates feature processing at a high dimension and improves the efficiency of feature processing.

[0072] Step S203: Divide the high-dimensional information features into N non-overlapping feature blocks. For any feature block, divide the feature block into M non-overlapping feature sub-blocks and determine the relative position of each feature sub-block in the feature block.

[0073] In this system, the number of channels in both feature blocks and feature sub-blocks is the same as the number of channels in the high-dimensional information features. N and M are both positive integers. The size of feature blocks is consistent, and the size of feature sub-blocks is consistent. The relative position can characterize the position information of a feature sub-block within its respective feature block.

[0074] Specifically, high-dimensional information features can be evenly divided during the unfolding process, which can be achieved using the unfold function.

[0075] The steps described above, which divide high-dimensional information features into N non-overlapping feature blocks, further divide any feature block into M non-overlapping feature sub-blocks, and determine the relative position of each feature sub-block within the feature block, divide the high-dimensional information features twice. This allows the feature blocks to encode global information based on the local information contained in the feature sub-blocks during subsequent encoding processes, thereby improving the representational ability of the subsequent encoding results.

[0076] Step S204: According to the preset query order, form a sub-block sequence by forming the feature sub-blocks that represent the same relative position in all feature blocks, and obtain the sub-block sequence of the corresponding relative position. Use an encoder to encode each sub-block sequence to obtain the encoding sequence of the corresponding sub-block sequence, and determine the encoding result of each feature sub-block from the corresponding encoding sequence.

[0077] The query order can represent the selection order of feature blocks, that is, the arrangement order of feature sub-blocks forming sub-block sequences in each feature block. A feature sequence can include N feature sub-blocks corresponding to the feature block. The encoding sequence can refer to the result after encoding the sub-block sequence. The encoding result can refer to the encoding result corresponding to the feature sub-block.

[0078] Specifically, the encoding process can be implemented using the Transformer model. A feature sub-block contains local information corresponding to its own feature block, while the sequence of sub-blocks contains global information of all feature blocks.

[0079] The above steps, following a preset query order, form a sub-block sequence from all feature blocks representing the same relative position, obtain the sub-block sequence of corresponding relative positions, encode each sub-block sequence using an encoder to obtain the encoding sequence of the corresponding sub-block sequence, and determine the encoding result of each feature sub-block from the corresponding encoding sequence. By equally dividing feature blocks and feature sub-blocks, the feature sub-blocks represent local information, and then global encoding is performed. This approach can better extract local and global information and improve the accuracy of feature position recovery, thus adapting to the characteristics of image segmentation tasks with strong constraints on pixel position.

[0080] Step S205: According to the relative positions of all feature sub-blocks, restore all encoding results to the encoding features of the corresponding feature blocks, use a pooling layer to perform pooling operation on the encoding features to obtain pooled features, perform feature decoding based on the pooled features to obtain the decoding result, and determine the decoding result as the tongue image segmentation result.

[0081] Among them, the encoded features can represent the feature information of the tongue image, the pooling layer can be used to aggregate the feature information of the encoded features, the pooling features can represent the result of the feature information aggregation, and the decoding result can refer to the result of the pooling feature decoding process.

[0082] Specifically, after obtaining the encoded sequence of the corresponding sub-block sequence, a folding operation is required to restore all encoded sequences to encoded features. The folding operation can be implemented using the fold function.

[0083] Optionally, after restoring all encoding results to the encoded features of the corresponding feature blocks, the process also includes:

[0084] The dimensionality of the encoded features is reduced by using a third-channel convolutional layer. The dimensionality-reduced result is then concatenated with the high-dimensional information features to obtain the concatenated result.

[0085] A fourth-channel convolutional layer is used to increase the dimensionality of the splicing result, and the increased dimensionality result is determined as the target feature.

[0086] Accordingly, pooling layers are used to perform pooling operations on the encoded features to obtain pooled features, including:

[0087] Pooling layers are used to perform pooling operations on the target features to obtain pooled features.

[0088] The third channel convolutional layer can be a 1*1 convolutional layer to project the encoded features to a low dimension. The third channel convolutional layer can be a 1*1 convolutional layer to increase the dimension of the concatenation result to the same dimension as the high-dimensional information feature. The dimension reduction result can refer to the projection result of the encoded features in the low dimension. The concatenation result can contain the shallow high-dimensional information feature and the deep dimension reduction result. The size of the target feature is consistent with the size of the high-dimensional information feature.

[0089] In this embodiment, the concept of residual connection is incorporated into the encoding process to ensure that the target features have good representation capabilities, avoid the situation where the target features' representation capabilities decline, and improve the accuracy of subsequent image segmentation based on target features.

[0090] Optionally, the pooling layer includes an average pooling layer, a pooling channel convolutional layer, and K dilated convolutional layers. The dilation rate of any two dilated convolutional layers is different. The dilation rate is used to control the receptive field of the corresponding dilated convolutional layer, and K is an integer greater than one.

[0091] Pooling layers are used to perform pooling operations on the encoded features to obtain pooled features, including:

[0092] The encoded features are input into the mean pooling layer, the pooling channel convolutional layer, and all dilated convolutional layers respectively to obtain the mean pooling result, the pooling channel convolution result, and the K dilated convolution results;

[0093] The mean pooling result is upsampled to obtain the upsampled result. The upsampled result, the pooling channel convolution result, and the K dilated convolution results are fused to obtain the fused result, which is used as the pooling feature.

[0094] Among them, the mean pooling layer can refer to the adaptive mean pooling layer, the pooling channel convolutional layer can be implemented using a 1*1 convolutional layer, and one dilated convolutional layer corresponds to one dilation rate.

[0095] Specifically, the encoded features are input into the mean pooling layer for pooling processing. The pooling result can be processed by a preset 1*1 convolution kernel and then upsampled to obtain the mean pooling result. In this embodiment, the upsampling can be done using bilinear interpolation. In this embodiment, K can be 3, which means it contains three dilated convolution layers.

[0096] Optionally, after obtaining the fusion result, the following may also be included:

[0097] The fusion result is processed using a global average pooling layer to obtain a global average pooling result. The size of the global average pooling result is reconstructed, and a one-dimensional convolution kernel is used to perform a convolution operation on the reconstructed result to obtain a one-dimensional convolution result.

[0098] The one-dimensional convolution result is activated using an activation function to obtain the activation result. The size of the activation result is then reconstructed to obtain the target reconstruction result.

[0099] Multiply the target reconstruction result and the fusion result to obtain the multiplied result;

[0100] Accordingly, the pooling features used as the fusion results include:

[0101] The result of multiplication is used as the pooling feature.

[0102] Among them, the global average pooling layer can be used to integrate global feature information, the global average pooling result can be used to represent global feature information, the one-dimensional convolution kernel can be used to realize local cross-channel interaction within the feature, the one-dimensional convolution result can refer to the result after the reconstruction result is convolved in one dimension, and the activation function can be the sigmoid function.

[0103] In this embodiment, global average pooling can effectively reduce the number of parameters and integrate global feature information. Furthermore, the use of one-dimensional convolution, compared to the conventional two-dimensional convolution, can achieve local cross-channel feature information interaction and avoid the poor feature extraction effect caused by dimensionality reduction operations. Thus, while ensuring network performance, the complexity of the network is reduced.

[0104] Optionally, feature decoding is performed based on pooling features to obtain the decoding result, which is then determined to be the tongue image segmentation result, including:

[0105] The pooling features are convolved using an encoding convolutional layer to obtain the encoding convolution result.

[0106] The encoding convolution result is interpolated using bilinear interpolation. The interpolated result is then used as the decoding result, which in turn is used as the tongue segmentation result.

[0107] Among them, the encoding convolutional layer can be used to further extract the feature information of pooling features, that is, the encoding convolution result. The bilinear interpolation method can be used for upsampling so that the size of the interpolation result is consistent with the tongue image. The tongue image segmentation result can refer to the semantic segmentation map of the tongue image. In this embodiment, in the semantic segmentation map, the pixel value of the pixel belonging to the tongue image is 1, and the pixel value of other pixels is 0.

[0108] In this embodiment, bilinear interpolation is used to generate semantic segmentation maps. This method is simple to calculate and has good interpolation results, thereby improving the accuracy of tongue image segmentation.

[0109] The above steps involve restoring all encoding results to the corresponding encoding features of the feature blocks according to their relative positions, using a pooling layer to perform pooling operations on the encoding features to obtain pooled features, performing feature decoding based on the pooled features to obtain the decoding result, and determining the decoding result as the tongue image segmentation result. Further pooling and decoding are performed based on the encoding features that can represent local and global features, thereby obtaining a decoding result that fully utilizes feature information and thus improving the accuracy of tongue image segmentation.

[0110] In this embodiment, local information features are extracted through convolution processing, and then the features are segmented so that each feature sub-block contains the local information of its respective feature block. Then, the sub-block sequence at the same position is encoded by an encoder to extract global information. This makes the encoded features contain both local and global information, which improves the richness and representation ability of the encoded features, thereby improving the effect of tongue image segmentation.

[0111] Corresponding to the artificial intelligence-based tongue image extraction method in the above embodiment, Figure 3 The diagram illustrates a structural block diagram of an artificial intelligence-based tongue image extraction device according to Embodiment 2 of the present invention. This tongue image extraction device is applied to a server, and the corresponding computer device is connected to the client to obtain the tongue image to be segmented. The server-side computer device is equipped with a feature extraction layer and a pooling layer. The feature extraction layer is used to extract image features from the tongue image, and the pooling layer is used to perform dimensionality reduction processing on the input features to aggregate feature information. For ease of explanation, only the parts relevant to the embodiments of the present invention are shown.

[0112] See Figure 3 The tongue image extraction device includes:

[0113] Feature extraction module 31 is used to input the tongue image to be segmented into the feature extraction layer for feature extraction to obtain initial image features;

[0114] The convolution processing module 32 is used to perform convolution processing on the initial image features using the first convolution kernel to obtain local information features, and to perform convolution processing on the local information features using the second convolution kernel to obtain high-dimensional information features. The high-dimensional information features and the local information features have the same feature size, and the number of channels of the high-dimensional information features is greater than the number of channels of the local information features.

[0115] The feature segmentation module 33 is used to divide the high-dimensional information feature into N non-overlapping feature blocks. For any feature block, the feature block is divided into M non-overlapping feature sub-blocks. The relative position of each feature sub-block in the feature block is determined. The number of channels of the feature block and the number of channels of the feature sub-block are the same as the number of channels of the high-dimensional information feature. N and M are both positive integers.

[0116] The sequence encoding module 34 is used to form a sub-block sequence by forming the feature sub-blocks that represent the same relative position in all feature blocks according to a preset query order, thereby obtaining the sub-block sequence of the corresponding relative position. The encoder is used to encode each sub-block sequence to obtain the encoding sequence of the corresponding sub-block sequence, and the encoding result of each feature sub-block is determined from the corresponding encoding sequence.

[0117] The tongue image segmentation module 35 is used to restore all encoding results to the encoding features of the corresponding feature blocks according to the relative positions of all feature sub-blocks, use a pooling layer to perform pooling operations on the encoding features to obtain pooled features, perform feature decoding based on the pooled features to obtain the decoding result, and determine the decoding result as the tongue image segmentation result.

[0118] Optionally, the feature extraction layer includes a first local convolutional layer, a first channel convolutional layer, and a batch normalization layer;

[0119] The aforementioned feature extraction module 31 includes:

[0120] The first parallel convolutional unit is used to input the tongue image into the first local convolutional layer and the first channel convolutional layer respectively, and both are processed by the batch normalization layer to obtain the first convolution result and the second convolution result;

[0121] The first convolutional fusion unit is used to add the first convolution result and the second convolution result, determine the added result as the first convolutional feature, and use the first convolutional feature as the initial image feature.

[0122] Optionally, the feature extraction layer may also include a second local convolutional layer and a second channel convolutional layer;

[0123] The feature extraction module 31 mentioned above also includes:

[0124] The second parallel convolutional unit is used to input the first convolutional features into the second local convolutional layer and the second channel convolutional layer respectively, and both are processed by the batch normalization layer to obtain the third convolutional result and the fourth convolutional result.

[0125] The second convolutional fusion unit is used to input the first convolutional features into the batch normalization layer for processing, obtain the normalization result, and add the third convolutional result, the fourth convolutional result and the normalization result to determine the added result as the second convolutional feature.

[0126] Accordingly, the first convolutional fusion unit mentioned above includes:

[0127] The feature determination subunit is used to use the second convolutional features as initial image features.

[0128] Optionally, the above-mentioned tongue image extraction device also includes:

[0129] The feature dimensionality reduction module is used to reduce the dimensionality of the encoded features using a third-channel convolutional layer to obtain the dimensionality reduction result. The dimensionality reduction result is then concatenated with the high-dimensional information features to obtain the concatenated result.

[0130] The feature upscaling module is used to upscale the splicing result using a fourth-channel convolutional layer to determine the upscaled result as the target feature.

[0131] Accordingly, the tongue image segmentation module 35 includes:

[0132] The feature pooling unit is used to perform pooling operations on the target features using pooling layers to obtain pooled features.

[0133] Optionally, the pooling layer includes an average pooling layer, a pooling channel convolutional layer, and K dilated convolutional layers. The dilation rate of any two dilated convolutional layers is different. The dilation rate is used to control the receptive field of the corresponding dilated convolutional layer, and K is an integer greater than one.

[0134] The above-mentioned tongue image segmentation module 35 includes:

[0135] Multi-layer pooling units are used to input the encoded features into the mean pooling layer, the pooling channel convolutional layer, and all dilated convolutional layers respectively, to obtain the mean pooling result, the pooling channel convolution result, and the K dilated convolution results;

[0136] The feature fusion unit is used to upsample the mean pooling result to obtain the upsampled result. The upsampled result, the pooling channel convolution result, and the K dilated convolution results are then fused to obtain the fused result, which is used as the pooling feature.

[0137] Optionally, the feature fusion unit mentioned above further includes:

[0138] The global pooling subunit is used to process the fusion result using the global average pooling layer to obtain the global average pooling result, to resize the global average pooling result, and to perform a convolution operation on the resized result using a one-dimensional convolution kernel to obtain a one-dimensional convolution result.

[0139] The size reconstruction subunit is used to activate the one-dimensional convolution result using an activation function to obtain the activation result, and then to reconstruct the size of the activation result to obtain the target reconstruction result.

[0140] The multiplication calculation subunit is used to multiply the target reconstruction result and the fusion result to obtain the multiplication result;

[0141] Pooling features determine sub-units, which are used to determine the pooling features based on the multiplication result.

[0142] Optionally, the tongue segmentation module 35 mentioned above includes:

[0143] The encoding processing unit is used to perform convolution processing on the pooling features using the encoding convolutional layer to obtain the encoding convolution result;

[0144] The interpolation processing unit is used to perform interpolation processing on the encoded convolution result using bilinear interpolation, determine the interpolation processing result as the decoding result, and determine the decoding result as the tongue image segmentation result.

[0145] It should be noted that the information interaction and execution process between the above modules, units, and sub-units are based on the same concept as the method embodiments of the present invention. For details on their specific functions and technical effects, please refer to the method embodiments section, which will not be repeated here.

[0146] Figure 4 This is a schematic diagram of the structure of a computer device provided in Embodiment 3 of the present invention. Figure 4 As shown, the computer device of this embodiment includes: at least one processor ( Figure 4 Only one is shown in the diagram), a memory, and a computer program stored in the memory and executable on at least one processor, which, when executed by the processor, implements the steps in any of the above-described tongue image extraction method embodiments.

[0147] This computer device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that... Figure 4 The examples of computer devices are merely examples and do not constitute a limitation on computer devices. Computer devices may include more or fewer components than shown in the illustration, or combinations of certain components, or different components, such as network interfaces, displays, and input devices.

[0148] The processor referred to can be a CPU, but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0149] Memory includes readable storage media, internal memory, etc., wherein internal memory can be the RAM of a computer device, providing an environment for the operation of the operating system and computer-readable instructions stored in the readable storage media. The readable storage media can be the hard drive of a computer device, or in other embodiments, it can be an external storage device of the computer device, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, memory can include both internal storage units and external storage devices of a computer device. Memory is used to store the operating system, applications, bootloader, data, and other programs, such as program code for computer programs. Memory can also be used to temporarily store data that has been output or will be output.

[0150] Those skilled in the art will understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the functions described above can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this invention. The specific working process of the units and modules in the above device can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here. If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention can implement all or part of the processes in the methods of the above embodiments by instructing related hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the above method embodiments. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. A computer-readable medium can include at least: any entity or device capable of carrying computer program code, a recording medium, a computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0151] The present invention can implement all or part of the processes in the methods of the above embodiments, or it can be accomplished by a computer program product. When the computer program product is run on a computer device, the computer device executes the steps in the above method embodiments.

[0152] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0153] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0154] In the embodiments provided by this invention, it should be understood that the disclosed apparatus / computer devices and methods can be implemented in other ways. For example, the apparatus / computer device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0155] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0156] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A tongue image extraction method based on artificial intelligence, characterized in that, The tongue image extraction method includes: The tongue image to be segmented is input into the feature extraction layer for feature extraction to obtain the initial image features; The initial image features are convolved using a first convolution kernel to obtain local information features. The local information features are then convolved using a second convolution kernel to obtain high-dimensional information features. The high-dimensional information features and the local information features have the same feature size, and the number of channels in the high-dimensional information features is greater than the number of channels in the local information features. The high-dimensional information feature is divided into N non-overlapping feature blocks. For any feature block, the feature block is divided into M non-overlapping feature sub-blocks. The relative position of each feature sub-block in the feature block is determined. The number of channels of the feature block and the number of channels of the feature sub-block are the same as the number of channels of the high-dimensional information feature. N and M are both positive integers. According to the preset query order, feature sub-blocks representing the same relative position in all feature blocks are formed into a sub-block sequence, and the sub-block sequence corresponding to the relative position is obtained. The encoder is used to encode each sub-block sequence to obtain the encoding sequence of the corresponding sub-block sequence. The encoding result of each feature sub-block is determined from the corresponding encoding sequence. Based on the relative positions of all feature sub-blocks, all encoding results are restored to the encoding features corresponding to the feature blocks. Pooling is performed on the encoding features to obtain pooled features. Feature decoding is performed based on the pooled features to obtain the decoding result. The decoding result is determined to be the tongue image segmentation result. After restoring all encoding results to the encoded features corresponding to the feature blocks, the method further includes: The dimensionality of the encoded features is reduced by using a third-channel convolutional layer to obtain the dimensionality reduction result. The dimensionality reduction result is then concatenated with the high-dimensional information features to obtain the concatenated result. The splicing result is up-dimensionalized using a fourth-channel convolutional layer, and the up-dimensionalized result is determined as the target feature. Accordingly, the process of using a pooling layer to perform a pooling operation on the encoded features to obtain pooled features includes: The pooling layer is used to perform a pooling operation on the target feature to obtain the pooled feature; The pooling layer includes an average pooling layer, a pooling channel convolutional layer, and K dilated convolutional layers. The dilation rates of any two dilated convolutional layers are different. The dilation rates are used to control the receptive field of the corresponding dilated convolutional layers. K is an integer greater than one. The process of using a pooling layer to perform a pooling operation on the encoded features to obtain pooled features includes: The encoded features are respectively input into the mean pooling layer, the pooling channel convolutional layer, and all dilated convolutional layers to obtain the mean pooling result, the pooling channel convolution result, and K dilated convolution results; The mean pooling result is upsampled to obtain an upsampled result. The upsampled result, the pooling channel convolution result, and the K dilated convolution results are fused to obtain a fused result. The fused result is used as the pooling feature. After obtaining the fusion result, the following is also included: The fusion result is processed using a global average pooling layer to obtain a global average pooling result. The size of the global average pooling result is reconstructed, and a one-dimensional convolution kernel is used to perform a convolution operation on the reconstructed result to obtain a one-dimensional convolution result. The one-dimensional convolution result is activated using an activation function to obtain an activation result. The size of the activation result is then reconstructed to obtain the target reconstruction result. Multiply the target reconstruction result and the fusion result to obtain the multiplication result; Accordingly, using the fusion result as the pooling feature includes: The multiplication result is used as the pooling feature.

2. The tongue image extraction method according to claim 1, characterized in that, The feature extraction layer includes a first local convolutional layer, a first channel convolutional layer, and a batch normalization layer; The process of inputting the tongue image to be segmented into the feature extraction layer for feature extraction to obtain initial image features includes: The tongue image is input into the first local convolutional layer and the first channel convolutional layer respectively, and both are processed by the batch normalization layer to obtain the first convolution result and the second convolution result; The first convolution result and the second convolution result are added together, and the sum is determined as the first convolution feature. The first convolution feature is then used as the initial image feature.

3. The tongue image extraction method according to claim 2, characterized in that, The feature extraction layer further includes a second local convolutional layer and a second channel convolutional layer; After determining that the summation result is the first convolutional feature, the following steps are also included: The first convolutional features are input into the second local convolutional layer and the second channel convolutional layer, respectively, and both are processed by the batch normalization layer to obtain the third convolutional result and the fourth convolutional result. The first convolutional feature is input into the batch normalization layer for processing to obtain a normalization result. The third convolutional result, the fourth convolutional result, and the normalization result are added together to determine the added result as the second convolutional feature. Accordingly, using the first convolutional feature as the initial image feature includes: The second convolutional feature is used as the initial image feature.

4. The tongue image extraction method according to any one of claims 1 to 3, characterized in that, The step of performing feature decoding based on the pooling features to obtain a decoding result, and determining the decoding result as a tongue image segmentation result, includes: The pooling features are convolved using an encoding convolutional layer to obtain the encoding convolution result. The encoded convolution result is interpolated using bilinear interpolation, and the interpolation result is determined to be the decoding result, which is then determined to be the tongue image segmentation result.

5. A tongue image extraction device based on artificial intelligence, characterized in that, The tongue image extraction device includes: The feature extraction module is used to input the tongue image to be segmented into the feature extraction layer for feature extraction to obtain the initial image features; The convolution processing module is used to perform convolution processing on the initial image features using a first convolution kernel to obtain local information features, and to perform convolution processing on the local information features using a second convolution kernel to obtain high-dimensional information features. The high-dimensional information features and the local information features have the same feature size, and the number of channels of the high-dimensional information features is greater than the number of channels of the local information features. The feature segmentation module is used to divide the high-dimensional information feature into N non-overlapping feature blocks. For any feature block, the feature block is divided into M non-overlapping feature sub-blocks. The relative position of each feature sub-block in the feature block is determined. The number of channels of the feature block and the number of channels of the feature sub-block are the same as the number of channels of the high-dimensional information feature. N and M are both positive integers. The sequence encoding module is used to form a sub-block sequence by forming the feature sub-blocks that represent the same relative position in all feature blocks according to a preset query order, thereby obtaining the sub-block sequence of the corresponding relative position. The encoder is used to encode each sub-block sequence to obtain the encoding sequence of the corresponding sub-block sequence, and the encoding result of each feature sub-block is determined from the corresponding encoding sequence. The tongue image segmentation module is used to restore all encoding results to the encoding features corresponding to the feature blocks according to the relative positions of all feature sub-blocks, perform pooling operation on the encoding features using a pooling layer to obtain pooled features, perform feature decoding based on the pooled features to obtain decoding results, and determine the decoding results as tongue image segmentation results; Also includes: The feature dimensionality reduction module is used to reduce the dimensionality of the encoded features by a third-channel convolutional layer after all the encoded results are restored to the encoded features corresponding to the feature blocks, to obtain the dimensionality reduction result, and to concatenate the dimensionality reduction result with the high-dimensional information features to obtain the concatenated result; The feature upscaling module is used to upscale the splicing result using a fourth-channel convolutional layer to determine the upscaled result as the target feature. Accordingly, the tongue image segmentation module includes: A feature pooling unit is used to perform a pooling operation on the target feature using the pooling layer to obtain the pooled feature; The pooling layer includes an average pooling layer, a pooling channel convolutional layer, and K dilated convolutional layers. The dilation rates of any two dilated convolutional layers are different. The dilation rates are used to control the receptive field of the corresponding dilated convolutional layers. K is an integer greater than one. The tongue image segmentation module includes: A multi-layer pooling unit is used to input the encoded features into the mean pooling layer, the pooling channel convolutional layer and all dilated convolutional layers respectively, to obtain the mean pooling result, the pooling channel convolution result and K dilated convolution results; The feature fusion unit is used to upsample the mean pooling result to obtain an upsampled result, and to fuse the upsampled result, the pooling channel convolution result and the K dilated convolution results to obtain a fusion result, and to use the fusion result as the pooling feature. The feature fusion unit further includes: The global pooling subunit is used to process the fusion result using a global average pooling layer after obtaining the fusion result, to obtain a global average pooling result, to resize the global average pooling result, and to perform a convolution operation on the resized result using a one-dimensional convolution kernel to obtain a one-dimensional convolution result. The size reconstruction subunit is used to activate the one-dimensional convolution result using an activation function to obtain an activation result, and then to reconstruct the size of the activation result to obtain the target reconstruction result. The multiplication calculation subunit is used to multiply the target reconstruction result and the fusion result to obtain the multiplication result; Accordingly, using the fusion result as the pooling feature includes: A pooling feature determination subunit is used to use the multiplication result as the pooling feature.

6. A computer device, characterized in that, The computer device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the tongue image extraction method as described in any one of claims 1 to 4.

7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the tongue image extraction method as described in any one of claims 1 to 4.

Images