Tooth malformation judgment method adopting ShuffleNet convolution model

Abstract

The invention relates to the technical field of image recognition, in particular to a tooth malformation judgment method adopting a ShuffleNet convolution model. Comprising the following steps: obtaining a to-be-detected tooth image, performing image preprocessing on the image, and measuring and calculating tooth similarity factors; inputting the preprocessed image into a tooth malformation judgment model, and sequentially passing through a feature extraction layer, a feature training layer and a judgment classification layer; the feature training layer comprises three structures which are used for training different features respectively; the judgment output layer is used for identifying and judging the data features and outputting a judgment result, and the judgment result is used for distinguishing whether the teeth are deformed or not. By thinking the characteristics of the teeth, the tooth malformation judgment model is provided, the tooth shape, the tooth gap and the tooth texture are identified, and the tooth similarity factor is introduced into the model, so that the judgment accuracy of the model on the tooth malformation is improved.

Description

Technical Field

[0001] This invention relates to the field of dental malformation identification and judgment technology, specifically a method for dental malformation judgment using the ShuffleNet convolutional model. Background Technology

[0002] Dental health is one of the foundations for a normal life and is closely related to our daily routines. If we don't pay attention to the protection and use of our teeth, we are prone to dental diseases that threaten our dental health. Common dental diseases, in addition to cavities and periodontitis, also include dental malformations.

[0003] Common dental malocclusions include crowded teeth, malocclusion, sparse teeth, deep overbite, and deep overjet. These malocclusions can range from mildly affecting the appearance of teeth and causing difficulty in biting to more severe leading to more serious oral diseases and seriously threatening our health. This is especially important for adolescents, who are in a crucial period of rapid tooth development. Timely detection and treatment of dental malocclusions are essential; otherwise, the malocclusions will worsen, affecting their normal learning and daily life.

[0004] In addition to hospital examinations, existing methods for detecting dental malformations can also be based on image recognition. Relatively speaking, image recognition saves time and effort and is more suitable as a preventative detection method.

[0005] Patent CN201910125415.4 discloses a method for identifying defective teeth based on deep learning. It utilizes a CIFAR10Net deep learning network model to perform transfer learning on a dental image dataset, training and optimizing the network structure and parameters. The resulting trained neural network model is then used to identify defective regions in new images to be judged. This invention applies deep learning to practical defective tooth image recognition, effectively identifying dental defects.

[0006] Patent CN201710494251.3 describes an automated classification and design method for malocclusion based on artificial intelligence. This method automatically identifies and selects dental arch feature points, establishes a reference system, and uses Markov decision-making combined with deep learning and reinforcement learning for dental arch design, thereby achieving automated classification and design of malocclusion without human intervention.

[0007] Patent CN201810180256.3 discloses an image-based teeth correction method, device, and terminal equipment; the method involves determining the tooth region in the face region based on the obtained color image; registering the tooth region with the depth image and obtaining the features of the teeth in the tooth region; comparing the features of the teeth with standard tooth features; and performing orthodontic processing on the teeth based on the standard tooth features when the difference between the features of the teeth and the standard tooth features is greater than a preset threshold.

[0008] The aforementioned methods for identifying dental malformations often neglect the characteristics of the tooth image and fail to delve deeply into the features of the teeth, resulting in relatively low accuracy in identifying dental malformations. To address this, this invention proposes a dental malformation identification method using the ShuffleNet convolutional model, which can accurately determine whether teeth are malformed. Summary of the Invention

[0009] To address the problem that existing technologies still lack accuracy in image recognition of dental malformations, this invention provides a method for identifying dental malformations using a ShuffleNet convolutional model. By establishing a dental malformation identification model, the method trains on tooth shape, gaps between teeth, and tooth texture, and introduces a tooth similarity factor to achieve accurate identification of dental malformations.

[0010] To achieve the above objectives, the present invention provides the following technical solution, including:

[0011] The teeth to be tested are photographed to obtain an image; the image to be tested is preprocessed to obtain a tooth feature image; and the tooth similarity factor of the tooth feature image is calculated.

[0012] The tooth feature image is input into the tooth deformity judgment model; the tooth feature image passes through the feature extraction layer, feature training layer and judgment classification layer in the tooth deformity judgment model in sequence; the tooth deformity judgment model outputs the classification result; the classification result is used to determine whether there is a deformity in the teeth in the image to be detected.

[0013] The feature extraction layer is used to extract tooth features from the tooth feature image, including tooth shape features, tooth gap features, and tooth texture features.

[0014] The feature training layer is used to learn and train the tooth feature image. It includes three structures: a first structure for recognizing tooth shape features, a second structure for recognizing tooth gap features, and a third structure for recognizing tooth texture features. The trained image features output by the three structures are fused to form a combined vector.

[0015] The judgment and classification layer is used to identify and classify the combined vector; the judgment and classification layer includes a tooth similarity factor for weighting the combined vector.

[0016] The calculation process for the tooth similarity factor is as follows: The image to be detected is converted into a grayscale image; the tooth portion of the grayscale image is identified and segmented using a CNN convolution algorithm to obtain a tooth grayscale image; the tooth grayscale image is divided into two equal-sized left and right parts, and the left-right symmetry similarity s1 is calculated; the tooth grayscale image is divided into two equal-sized upper and lower parts, and the upper and lower symmetry similarity s2 is calculated; the tooth similarity factor is then calculated based on the tooth similarity function.

[0017]

[0018] Where e is the natural constant; s1 is the left-right symmetry similarity; s2 is the top-bottom symmetry similarity; and s is the derived tooth similarity factor, which is used as a weighted factor in the model to determine tooth deformity. Based on the approximate symmetry of teeth, the tooth similarity factor can, to some extent, reflect the degree of tooth deformity.

[0019] The image preprocessing process uses a homomorphic filtering algorithm to optimize the image; the homomorphic filtering algorithm includes a combined Butterworth filter:

[0020]

[0021] Where (u,v) are the coordinates of the point; D(u,v) is the distance of the point (u,v) from the origin of the frequency; D0 is the cutoff frequency, which is the function value D(0,0) corresponding to the point (0,0); n is the order; H(u,v) is the output value; p and θ are adjustment parameters.

[0022] The first structure includes a first grouped convolutional layer, a channel recombination layer, a first DepthWise convolutional layer, a second grouped convolutional layer, an average pooling layer, a feature combination layer, a dual attention mechanism module, and a convolutional layer.

[0023] The image features are input into the first structure, processed by an average pooling layer, and output as dataset U0; the image features are processed by the first grouped convolutional layer, and output as dataset U1; dataset U1 is input into the channel reconstruction layer, and output as dataset U2; dataset U2 is input into the first DepthWise convolutional layer, and output as dataset U3; dataset U3 is input into the second grouped convolutional layer, and output as dataset U4; dataset U4 and dataset U0 are combined through a feature combination layer to obtain dataset U5; dataset U5 is processed by a dual attention mechanism module, and output as dataset U6; dataset U6 is processed by the convolutional layer, and output as dataset U7.

[0024] The second structure includes a third grouped convolutional layer, a channel reconstruction layer, a second DepthWise convolutional layer, a fourth grouped convolutional layer, a feature addition layer, a dual attention mechanism module, and a convolutional layer. After the image features are input into the second structure, they are processed by the third grouped convolutional layer to output dataset V1. Dataset V1 is then input into the channel reconstruction layer to output dataset V2. Dataset V2 is then input into the second DepthWise convolutional layer to output dataset V3. Dataset V3 is then input into the fourth grouped convolutional layer to output dataset V4. Dataset V4 and dataset V1 are added together through the feature addition layer to obtain dataset V5. Dataset V5 is then processed by the dual attention mechanism module to output dataset V6. Dataset V6 is then processed by the convolutional layer to output dataset V7.

[0025] The third structure is composed of the first structure and the second structure; the first structure processes the input image features and outputs preprocessed features to the second structure; the second structure trains the input preprocessed features and outputs processed features.

[0026] The training process of the dental malformation judgment model is as follows: collect dental images and construct a dental image dataset; perform image preprocessing on the image dataset; calculate the dental similarity factor of the image dataset; divide the preprocessed image dataset into a training set and a validation set; train the dental malformation judgment model using the training set; optimize the dental malformation judgment model using the validation set; and obtain the dental malformation judgment model.

[0027] The image dataset contains two labels: those indicating the presence of dental malformations and those indicating the absence of dental malformations.

[0028] The dental malformation assessment model includes a dual attention mechanism module; the dual attention mechanism module includes a spatial attention mechanism module and a channel attention mechanism module.

[0029] The dental malformation judgment model includes a batch normalization module; the batch normalization module is used to unify the feature data of the dental feature images to the same dimension; the batch normalization method within the batch normalization module is as follows:

[0030]

[0031]

[0032] Where, x i For input tooth feature data; μ β σ is the mean of the data. β 2 The variance of the data; The value is the normalized value; θ is a local minimum added to avoid the denominator being zero; y i This is the result after transformation and offset; α is the scale factor; β is the translation factor.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] 1. An improved version of the classic Butterworth filter, a combined Butterworth filter, is proposed. This combined Butterworth filter integrates the high-frequency and low-frequency channel functions of the classic Butterworth filter. By setting adjustable parameters, an improved function is obtained, with the high-frequency channel function as the primary component and the low-frequency channel function as a secondary component. In practical image enhancement, it can process image information from both high-frequency and low-frequency channels. The combined Butterworth filter effectively improves the optimization processing effect of uneven lighting in images, which helps in the accurate judgment of dental malformation models.

[0035] 2. This invention proposes a tooth similarity function and a tooth similarity factor. The tooth similarity function can calculate the tooth similarity factor based on the left-right and top-bottom similarity of the tooth to be tested. In a healthy state, the shape of teeth is approximately symmetrical. The tooth similarity factor can reflect the health status of the tooth to be tested to a certain extent. Therefore, the tooth similarity factor is used as a weight and important basis, influencing subsequent judgments of dental deformities. The tooth similarity factor can fully utilize the characteristics of teeth to improve the accuracy of judgments.

[0036] 3. This invention proposes a dental malformation judgment model, which is based on the ShuffleNet module and proposes three independent structures: a first structure for recognizing tooth shape features; a second structure for recognizing tooth gap features; and a third structure for recognizing tooth texture features. Multiple training results are combined to obtain the classification result. This invention improves the ShuffleNet model based on the characteristics of different tooth image features, separately training it to recognize tooth shape features, tooth gap features, and tooth texture features in tooth images. After feature fusion of the recognition training results, the model is weighted using tooth similarity factors and then classified, which can effectively and accurately determine whether teeth are malformed. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of a tooth image according to the present invention;

[0038] Figure 2 These are experimental figures comparing the effects of the image preprocessing method of the present invention.

[0039] Figure 3This is a comparative experimental diagram showing the effects of the tooth similarity factor of the present invention;

[0040] Figure 4 This is a schematic diagram of the training process of the dental malformation judgment model of the present invention;

[0041] Figure 5 This is a schematic diagram of the dental deformity assessment model of the present invention;

[0042] Figure 6 This is a schematic diagram of the first structure of the present invention;

[0043] Figure 7 This is a schematic diagram of the second structure of the present invention;

[0044] Figure 8 This is a schematic diagram of the judgment logic of the present invention. Detailed Implementation

[0045] The proposed method for judging dental malformations will be further explained below according to the accompanying drawings provided in this invention. The technical explanations and experimental data involved in the specific implementation section are merely a specific embodiment of this technical solution. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0046] Healthy teeth are fundamental to a healthy life and are closely related to our daily diet. Besides common dental caries and periodontitis, more serious dental deformities threaten dental health. Many people are not very aware of dental deformities, only paying attention when they become noticeable. In fact, dental deformities are very common in our daily lives. Besides the well-known protruding teeth, there are many other types, which can be summarized as follows: crowded teeth, malocclusion, sparse teeth, deep overbite, and deep overjet.

[0047] Crowded teeth are usually caused by an insufficiently small dental arch or a narrow alveolar bone, resulting in insufficient space for the teeth to grow out of place. Many hard-to-reach areas in a crowded and misaligned dental arch are difficult to clean, making it easy for plaque and tartar to accumulate and cause cavities.

[0048] Asymmetric jaw, mainly refers to the asymmetry between the left and right sides of the lower jaw, and sometimes it may also present as a deviated chin. In severe cases, it can cause serious asymmetry of the bilateral face, affecting appearance, chewing and pronunciation.

[0049] Sparse teeth, an oversized dental arch, teeth that are too small or too few, congenital tooth loss or tooth loss, all result in teeth that are too loosely spaced and scattered, affecting normal eating.

[0050] Deep overbite, commonly known as buck teeth, is a condition where the maximum horizontal distance from the incisal edge of the upper anterior teeth to the labial surface of the lower anterior teeth exceeds 3mm when biting down, and the upper teeth protrude significantly; this is a relatively common malformation in clinical practice.

[0051] Deep overbite is characterized by the upper teeth covering too much of the lower teeth when clenching, affecting aesthetics. Unlike buck teeth, the upper teeth in a deep overbite do not protrude significantly. In some severe cases of deep overbite, only the upper front teeth are visible from the front when clenching, while the lower front teeth are completely covered, and the lower front teeth often bite onto the palatal mucosa of the maxilla. This is a serious dentofacial deformity that must be corrected promptly.

[0052] Dental malocclusion can affect appearance, leading to a tendency to cover one's mouth, lack of confidence when smiling, and reluctance to take profile photos. It can also cause health problems, such as occlusion, pronunciation, chewing and swallowing issues, and may even lead to organic lesions.

[0053] The purpose of this invention is to provide a method for identifying dental malformations using a ShuffleNet convolutional model, which achieves the identification of dental malformations through image recognition technology. Please refer to [link / reference]. Figures 1 to 8 The technical solution is as follows:

[0054] Step 1: Take a picture of the tooth to be tested to obtain the image.

[0055] When taking photos of the teeth to be examined, the entire tooth needs to be shown in the image, such as... Figure 1 As shown.

[0056] Step 2: Perform image preprocessing on the image to be detected to obtain a tooth feature image.

[0057] When photographing teeth, the lighting environment inside the oral cavity is dim, requiring the use of external light sources to increase brightness. This can easily lead to uneven lighting in the tooth images, resulting in variations in brightness and color depth, which affects subsequent feature recognition. Therefore, image preprocessing is necessary.

[0058] In the image preprocessing stage, considering the impact of uneven illumination on the image, a homomorphic filtering algorithm is used to process the image. The homomorphic filtering algorithm can reduce the negative effects of illumination changes by reducing low-frequency information and increasing high-frequency information in the image, thereby sharpening the edges and details of the image, enhancing image contrast, and compressing the image brightness range.

[0059] The image f(x,y) can be represented by the product of the incident component j(x,y) and the reflected component h(x,y):

[0060] f(x,y)=j(x,y)*h(x,y)

[0061] The uneven illumination primarily affects the incident component j(x,y), which is a low-frequency signal; the reflected component h(x,y) is mainly a high-frequency signal, reflecting many details of the image. Therefore, processing uneven illumination in an image using a homomorphic filtering algorithm aims to minimize the incident component and enhance the reflected component.

[0062] The image processing steps using a homomorphic filtering algorithm are as follows:

[0063] The image function is represented by the product of the incident and reflected components, resulting in an image function expression. The logarithm of the image function expression is then taken to obtain a logarithmic expression. A Fourier transform is performed on the logarithmic expression to obtain a frequency domain expression. The frequency domain expression is processed using a filter function and inversely transformed to the spatial domain. An exponential operation is then performed to obtain the transformed image function.

[0064] The filter function used is a combined Butterworth filter:

[0065]

[0066] Where (u,v) are the coordinates of the point; D(u,v) is the distance of the point (u,v) from the origin of the frequency; D0 is the cutoff frequency, which is the function value D(0,0) corresponding to the point (0,0); n is the order; H(u,v) is the output value; p and θ are adjustment parameters.

[0067] The combined Butterworth filter is an improvement on the classic Butterworth filter. The classic Butterworth filter consists of two parts: a high-frequency channel function and a low-frequency channel function. In image preprocessing, the high-frequency channel function should be used to enhance the reflection component of the image. However, since the distinction between high and low frequencies is not absolute, useful image information can also exist in the low-frequency channel. Therefore, this invention proposes a combined Butterworth filter, which combines the high-frequency channel function and the low-frequency channel function with weights and sets adjustable parameters to increase the proportion of the high-frequency channel function, so as to accurately achieve the purpose of image enhancement.

[0068] To verify the actual effectiveness of the combined Butterworth filter, this invention compares different image preprocessing methods. This invention employs a controlled variable method, using models for dental malformation assessment, with the only difference being the image preprocessing method within each model.

[0069] The experimental group consisted of the first model using the combined Butterworth filter homomorphic filtering algorithm, while the control group consisted of three groups: the second model using the Butterworth filter homomorphic filtering algorithm; the third model using Gaussian filtering; and the fourth model without image preprocessing.

[0070] Using the collected image dataset, after image preprocessing, 70% of the image dataset was selected for each experiment by random sampling.

[0071] The first, second, third, and fourth models were trained using the training set to obtain the parameters for each model. The corresponding models were then tested using the test set to obtain the accuracy rate of the dental malformation judgment model. The accuracy rate is the ratio of the number of times the dental malformation judgment model correctly judged the malformation to the total number of judgments in the test set data. This experimental process was repeated 10 times, and the accuracy rate of each model was recorded.

[0072] The final experimental results are as follows Figure 2 As shown in the figure, the horizontal axis represents the number of experiments, and the vertical axis represents the recognition accuracy of the model. Different curves represent different experimental models.

[0073] Overall, the experimental results show that the first model using the combined Butterworth filter homomorphic filtering algorithm has the highest average accuracy of 92.91%, followed by the second model using the Butterworth filter homomorphic filtering algorithm with an accuracy of 87.71%, then the third model using Gaussian filtering with an accuracy of 87.05%, and finally the fourth model without image preprocessing with an accuracy of 82.42%.

[0074] Step 3: Calculate the tooth similarity factor of the tooth feature image.

[0075] Human teeth are approximately regular and symmetrical in a healthy state. However, when teeth become deformed, this characteristic is disrupted. Therefore, this invention measures the regularity and symmetry of teeth and uses it as an influencing factor in determining whether teeth are deformed. Based on this, this paper proposes a tooth similarity function and a tooth similarity factor.

[0076] The calculation process for the tooth similarity factor is as follows:

[0077] Convert the image to be detected into a grayscale image;

[0078] The tooth portion of the grayscale image is identified and segmented using a CNN convolution algorithm to obtain a grayscale image of the teeth.

[0079] The grayscale image of the teeth is divided into two equal parts, left and right, and the left-right symmetry similarity is calculated.

[0080] The grayscale image of the teeth is divided into two equal parts, upper and lower, and the symmetry similarity between the upper and lower parts is calculated.

[0081] The tooth similarity factor is calculated based on the tooth similarity function.

[0082]

[0083] Where e is the natural constant; s1 is the left-right symmetry similarity; s2 is the top-bottom symmetry similarity; and s is the derived tooth similarity factor, which is used as an influencing factor in judging tooth deformity and is used to weight the model.

[0084] To verify the actual effect of the tooth similarity factor in the model, a comparative study was conducted on the effects of using the tooth similarity factor and not using it. The controlled variable method was used, and the models used were all tooth malformation judgment models.

[0085] The experimental group was the first model that used the tooth similarity factor, and the control group was the second model that did not use the tooth similarity factor; the first model and the second model differed only in whether or not the tooth similarity factor was used.

[0086] Using the collected image dataset, after image preprocessing, 70% of the image dataset was selected for each experiment by random sampling.

[0087] The first and second models were trained using the training set to obtain the parameters of each model. The corresponding models were then tested using the test set to obtain the accuracy rate of the dental malformation judgment model. The accuracy rate is the ratio of the number of times the dental malformation judgment model correctly judged the malformation to the total number of judgments in the test set data. The above experimental process was repeated 10 times, and the accuracy rate of each model was recorded.

[0088] The final experimental results are as follows Figure 3 As shown in the figure, the horizontal axis represents the number of experiments conducted, and the vertical axis represents the model's recognition accuracy.

[0089] Overall, the experimental results show that the first model using the tooth similarity factor has a higher average accuracy of 92.50%, while the second model without the tooth similarity factor has an accuracy of 86.73%.

[0090] Step four: Identify and judge the tooth feature images using a tooth deformity judgment model.

[0091] This invention proposes a dental malformation judgment model to identify and judge dental images. The training process of the model is as follows: Figure 4 As shown.

[0092] The training process for the dental malformation assessment model is as follows:

[0093] Collect tooth images to construct a tooth image dataset; preprocess the image dataset; calculate the tooth similarity factor of the image dataset; divide the preprocessed image dataset into a training set and a validation set; train the tooth malformation judgment model using the training set; optimize the tooth malformation judgment model using the validation set; obtain the tooth malformation judgment model. The structure of the tooth malformation judgment model is as follows: Figure 5 As shown.

[0094] The tooth feature image is input into the tooth deformity judgment model; the tooth feature image passes through the feature extraction layer, feature training layer and judgment classification layer in the tooth deformity judgment model in sequence; the tooth deformity judgment model outputs the classification result; the classification result is used to determine whether there is a deformity in the teeth in the image to be detected.

[0095] The feature extraction layer is used to extract tooth features from the tooth feature image, including tooth shape features, tooth gap features, and tooth texture features.

[0096] The tooth shape features are used to identify the shape of teeth; when teeth are deformed, the shape of individual teeth and the overall shape of the teeth may change; by identifying and analyzing the tooth shape features, the accuracy of tooth deformity identification can be improved.

[0097] The tooth gap feature is used to identify abnormalities in tooth gaps in dental images. Normal tooth images are regular and fine, with the entire tooth area basically within the same color range. Dental malformations cause teeth to no longer be fine and tight, resulting in many gaps. This is reflected in dental images by a significant change in the color of the gaps compared to the surrounding teeth, resulting in irregular areas of varying colors. Therefore, the size and shape of the gaps can be identified by the color changes between the gaps, thereby exploring the condition of the teeth.

[0098] The tooth texture features are used to identify details in tooth images; when there are cases of overlapping upper and lower teeth or misalignment between teeth, the texture of the intersecting areas of the teeth can be used for identification.

[0099] The feature training layer is used to learn and train the tooth feature image. It includes three training structures: a first structure for recognizing tooth shape features, a second structure for recognizing tooth gap features, and a third structure for recognizing tooth texture features. The trained image features output by the three structures are fused to form a combined vector.

[0100] The classification layer is used to identify and classify the combined vectors; the classification layer includes a tooth similarity factor for weighting the combined vectors.

[0101] The first structure is derived from the ShuffleNet model and includes a first grouped convolutional layer, a channel recombination layer, a first DepthWise convolutional layer, a second grouped convolutional layer, an average pooling layer, a feature combination layer, a dual attention mechanism module, and a convolutional layer. Skip connections are used to combine the initial data with the processed data to avoid model overfitting and speed up model training.

[0102] The specific data processing procedure is as follows: the input feature map is first subjected to 1*1 grouped convolution, the resulting convolution is normalized, and then transformed using an activation function. This reduces the complexity of the network while performing dimensionality reduction. Considering that in 1*1 grouped convolution, feature data is only processed within a group, and the image features between different groups are not interconnected, the network's performance in feature extraction will be singular and independent, failing to effectively extract overall features. Therefore, channel recombination is employed to enhance the information fusion capability between channels.

[0103] The extracted image features are input into the first DepthWise convolutional layer, and the convolution result is input into the normalization layer to improve the training efficiency of the model. This operation is performed in the spatial dimension, with a single convolutional kernel responsible for the operation of a single channel. Next, 1*1 grouped convolutions are used again to restore the channel dimension to match the shortcut path while combining the feature maps to generate a new feature map. Then, it is fed into the normalization layer, and the output feature map is stacked with the feature map generated by the average pooling layer with a stride of 2 on the shortcut path. This can expand the feature dimension without increasing the computational cost. Finally, the feature map is output after passing through the ReLU activation function. The structure of the first structure is as follows: Figure 6 As shown.

[0104] The second structure is similar to the first structure, except that it does not use an average pooling layer in the skip connections, and it replaces the fusion of feature maps and positional features with addition. The second structure includes a third grouped convolutional layer, a channel recombination layer, a second DepthWise convolutional layer, a fourth grouped convolutional layer, a feature addition layer, a dual attention mechanism module, and convolutional layers. The structural composition of the second structure is as follows: Figure 7 As shown.

[0105] The third structure is a combination of the first structure and the second structure; the first structure processes the input image features and outputs preprocessed features to the second structure; the second structure trains the input preprocessed features and outputs processed features.

[0106] The dental malformation judgment model improves the ShuffleNet model based on the characteristics of different image features of teeth, and obtains a targeted training structure. It can be trained to recognize shape features, gap features and texture features separately. After feature fusion of the recognition training results, it is weighted by tooth similarity factors and classified, which can effectively determine whether there is a malformation of the teeth.

[0107] The grouped convolution divides the original data features into several groups, performs convolution within each group, and further enhances information flow between groups by grouping the results of each group convolution into subgroups within the same group and then exchanging subgroups within each group. Traditional 1x1 convolution consumes a lot of resources, while grouped convolution effectively alleviates this problem.

[0108] The batch normalization layer unifies the data to the same dimension. When training a neural network, the input data undergoes various convolutional processes, linear transformations, and nonlinear transformations, which may alter the data distribution. This change in data distribution increases with the depth of the network model, so normalization is necessary in the intermediate layers of the network. The specific steps are as follows:

[0109] First, calculate the mean of the data;

[0110]

[0111] Calculate the variance of the data;

[0112]

[0113] Normalization is performed; θ is a local minimum value added to avoid the denominator being zero.

[0114]

[0115] Perform scaling and offset transformations.

[0116]

[0117] Where, x i The input data is n; the number of input data is μ. β σ is the mean of the data. β 2 The variance of the data; The value is the normalized value; θ is a local minimum added to avoid the denominator being zero; y i This is the result after transformation and offset; α is the scale factor; β is the translation factor.

[0118] The channel recombination layer first converts the channel into a one-dimensional vector, processes it into a matrix according to a set format, performs a transpose operation on the matrix, and then expands the transposed matrix back into a one-dimensional vector.

[0119] The Depthwise convolution only changes the size of the feature map, not the number of channels; each kernel of the Depthwise convolution is responsible for only one channel, and each kernel convolves with only one channel; the number of kernels is equal to the number of input channels, and the number of output channels also remains unchanged, equal to the number of input channels, which is equal to the number of kernels.

[0120] The dual attention mechanism module can be viewed as a combination of three parts: the first part is the original data features and the channel attention module, which outputs processed data that has been weighted by channel attention; the second part is the original data features and the spatial attention module, which outputs processed data that has been weighted by spatial attention; and the third part is the original data. The above three parts of data are combined to obtain the final output data.

[0121] The channel attention module first aggregates the feature maps of each channel through a global average pooling layer, followed by two fully connected layers to obtain the weights of each channel, and finally, after normalization, obtains the weight matrix.

[0122] The spatial attention module first reduces the dimensionality of the original data features using a 1x1 convolutional layer, processing them into an appropriate format. Then, it optimizes the features using two 3x3 dilated convolutional layers. Finally, it generates weight parameters using another 1x1 convolutional layer. By combining spatial attention and channel attention mechanisms on the original data, it effectively preserves the original information, eliminates information redundancy, and improves the model's operating efficiency and recognition accuracy.

[0123] Step 5: Output the judgment result.

[0124] The above is a specific embodiment based on the technical solution of the present invention. Various changes made to the embodiment without departing from the basic principles and ideas of the present invention are within the protection scope of the present invention.

Claims

1. A method for determining dental malformations using a ShuffleNet convolutional model, characterized in that, This includes a dental malformation assessment model to determine whether teeth are deformed. The assessment process is as follows: The teeth to be examined are photographed to obtain an image; the image to be examined is then preprocessed to obtain a tooth feature image; Calculate the tooth similarity factor of the tooth feature image; The tooth feature image is input into the tooth deformity judgment model; the tooth feature image passes sequentially through the feature extraction layer, feature training layer, and judgment classification layer within the tooth deformity judgment model; the tooth deformity judgment model outputs the classification result; The classification results are used to determine whether there are malformations in the teeth in the image to be detected; The feature extraction layer is used to extract tooth features from the tooth feature image, including tooth shape features, tooth gap features, and tooth texture features; The feature training layer is used to learn and train the tooth feature image. It includes three structures: a first structure for recognizing tooth shape features, a second structure for recognizing tooth gap features, and a third structure for recognizing tooth texture features. The trained image features output by the three structures are fused to form a combined vector. The classification layer is used to identify and classify the combined vector; the classification layer includes a tooth similarity factor for weighting the combined vector.

2. The method for determining dental malformation using a ShuffleNet convolutional model according to claim 1, characterized in that: The calculation process for the tooth similarity factor is as follows: Convert the image to be detected into a grayscale image; The tooth portion of the grayscale image is identified and segmented using a CNN convolution algorithm to obtain a grayscale image of the teeth. The grayscale image of the teeth is divided into two equal parts, left and right, and the left-right symmetry similarity s1 is calculated. The grayscale image of the teeth is divided into two equal parts, upper and lower, and the symmetry similarity s2 between the upper and lower parts is calculated. The tooth similarity factor is calculated based on the tooth similarity function. Where e is the natural constant; s1 is the left-right symmetry similarity; s2 is the top-bottom symmetry similarity; and s is the derived tooth similarity factor, which is used as an influencing factor in judging tooth deformity and is used to weight the model.

3. The method for determining dental malformation using a ShuffleNet convolutional model according to claim 1, characterized in that: The image preprocessing process uses a homomorphic filtering algorithm to optimize the image; The homomorphic filtering algorithm includes a combined Butterworth filter: Where (u,v) are the coordinates of the point; D(u,v) is the distance of the point (u,v) from the origin of the frequency; D0 is the cutoff frequency, which is the function value D(0,0) corresponding to the point (0,0); n is the order; H(u,v) is the output value; p and θ are adjustment parameters.

4. The method for determining dental malformation using a ShuffleNet convolutional model according to claim 1, characterized in that: The first structure includes a first grouped convolutional layer, a channel recombination layer, a first DepthWise convolutional layer, a second grouped convolutional layer, an average pooling layer, a feature combination layer, a dual attention mechanism module, and a convolutional layer; The image features are input into the first structure, processed by the average pooling layer, and the output dataset U0 is obtained; the image features are processed by the first grouped convolutional layer, and the output dataset U1 is obtained. The dataset U1 is input into the channel reconstruction layer for processing, and the dataset U2 is output. The dataset U2 is input into the first DepthWise convolutional layer for processing, and the output dataset U3 is generated. The dataset U3 is input into the second grouped convolutional layer for processing, and the output dataset U4 is generated. The dataset U4 and the dataset U0 are combined using the feature combination layer to obtain dataset U5; The dataset U5 is processed by the dual attention mechanism module to output dataset U6; The dataset U6 is processed through the convolutional layer to output dataset U7.

5. The method for determining dental malformation using a ShuffleNet convolutional model according to claim 1, characterized in that: The second structure includes a third grouped convolutional layer, a channel recombination layer, a second DepthWise convolutional layer, a fourth grouped convolutional layer, a feature addition layer, a dual attention mechanism module, and convolutional layers; The image features are input into the second structure, processed by the third grouped convolutional layer, and the output dataset V1 is generated. The dataset V1 is input into the channel reconstruction layer for processing, and the dataset V2 is output. The dataset V2 is input into the second DepthWise convolutional layer for processing, and the output dataset V3 is generated. The dataset V3 is input into the fourth grouped convolutional layer for processing, and the output dataset V4 is generated. The dataset V4 and the dataset V1 are combined using the feature addition layer to obtain dataset V5; The dataset V5 is processed by the dual attention mechanism module to output dataset V6; The dataset V6 is processed through the convolutional layer to output dataset V7.

6. The method for determining dental malformation using a ShuffleNet convolutional model according to claim 1, characterized in that: The third structure is composed of the first structure and the second structure; The first structure processes the input image features and outputs the preprocessed features to the second structure; The second structure trains the input preprocessed features and outputs the processed features.

7. The method for determining dental malformation using a ShuffleNet convolutional model according to claim 1, characterized in that: The training process for the dental malformation assessment model is as follows: Collect tooth images and construct a tooth image dataset; Perform image preprocessing on the image dataset; Calculate the tooth similarity factor of the image dataset; The preprocessed image dataset is divided into a training set and a validation set; The dental deformity assessment model was trained using a training set. The dental malformation assessment model was optimized by using a validation set. The dental deformity assessment model was obtained.

8. The method for determining dental malformation using a ShuffleNet convolutional model according to claim 7, characterized in that: The image dataset contains two types of labels: They are distinguished as having dental malformations and not having dental malformations, respectively.

9. The method for determining dental malformation using a ShuffleNet convolutional model according to claim 1, characterized in that: The dental malformation assessment model includes a dual attention mechanism module; The dual attention mechanism module includes a spatial attention mechanism module and a channel attention mechanism module.

10. A method for determining dental malformation using a ShuffleNet convolutional model according to claim 1, characterized in that: The dental malformation judgment model includes a batch normalization module; the batch normalization module is used to unify the feature data of the dental feature images to the same dimension; the batch normalization method within the batch normalization module is as follows: Where, x i For input tooth feature data; μ β σ is the mean of the data. β 2 The variance of the data; The value is the normalized value; θ is a local minimum added to avoid the denominator being zero; y i This is the result after transformation and offset; α is the scale factor; β is the translation factor.

Images