A small sample-based high-precision tooth target detection method

Abstract

The present application relates to a kind of high-precision dental target detection method based on small sample.The present application uses the targeted amplification of biological characteristics in line with training image, improve the performance of dental segmentation model;For dental semantic edge image, dental segmentation image and the semantic segmentation image of the overlapping part between tooth, obtain high-precision dental semantic edge image;Propose "finite connected region" method, get single tooth profile image from dental semantic edge image by polar coordinate;Add the method based on position information to classify single tooth profile image, output the classification result of tooth.The present application uses four U-Net models to generate high-precision edge image by the targeted amplification of small sample under panoramic image and multi-region local image, forms high-precision dental semantic edge image, and only uses semantic edge image to extract the outline of single tooth.By the complementary advantages of tooth profile and position information, it achieves fast and accurate tooth classification.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of tooth image processing, and relates to a high-precision tooth target detection method based on a small sample. BACKGROUND

[0002] In today's high-tech information age, using pictures to store information has become an important technology. Computer image information processing combined with artificial intelligence can not only greatly improve the computer's ability to obtain and represent image information and knowledge generalization, but also can more conveniently and efficiently analyze computer images and detect targets.

[0003] Current target detection technology is relatively mature. RCNN is the first solution that can be truly applied in an industrial level. After that, Faster-RCNN follows RCNN and can quickly and accurately complete target detection after being trained by a large sample. However, the tooth target detection method still faces four problems. First, a large number of samples are needed to support, and it is difficult to achieve good results in a small sample situation. Second, the accuracy of semantic edge segmentation of teeth is low, and it is not easy to obtain a closed tooth segmentation edge. Third, the method for extracting a single tooth contour from a tooth overall image is not mature. Fourth, the contour difference between teeth is small, and it is difficult to obtain an accurate tooth classification result by using only the tooth contour.

[0004] The four problems faced by tooth target detection technology under small sample conditions generally have four tasks: task one is to deal with sample expansion under small sample conditions. Today's sample expansion is mostly random and direct repetition, enlargement, reduction, flip, and mirror image. Although it can quickly expand the sample, many medical images do not meet the biological characteristics after such operation, and the expanded images do not have the meaning of training samples. Task two is to accurately segment the dental image. This task has poor segmentation effect under small sample training. Today's technology for semantic segmentation is mostly through training of a large number of samples for the model, and through the model to generate a more accurate dental semantic segmentation image. However, the outline of each tooth on the dental X-ray image is low in clarity in the gum, resulting in inaccurate semantic edges of the dental image generated by the model. Task three is to extract the single tooth contour from the dental semantic image. There are various methods for extracting a single tooth contour from a dental semantic segmentation image, some directly through position information box to obtain a single tooth, and some extract a single tooth from a dental semantic image. However, the existing method for extracting a single tooth contour from a dental image generally combines a variety of existing technologies to maximize the accuracy, which is complex and time-consuming. Or directly recognize the image after semantic segmentation, which is a crude way, fast but with low accuracy. Task four is the training task of the dental target classification model, such as distinguishing different types of teeth or distinguishing upper and lower teeth. This task often corresponds to one label per image. The ordinary tooth classification method directly classifies the tooth through the tooth contour, but the difference between the teeth is small, and the ordinary classification method has low accuracy in distinguishing the types of teeth.

[0005] Chinese Patent Publication No. CN114219804A discloses a small sample tooth detection method based on a prototype segmentation network, which utilizes the respective advantages of target detection and semantic segmentation, reduces tooth segmentation type errors, and increases tooth boundary accuracy. In this technical solution, the prototype segmentation network is used, and the small sample tooth detection method based on the prototype segmentation network is improved to improve the accuracy of the tooth segmentation boundary. However, it still cannot guarantee the closure of the semantic edge. Chinese Patent Publication No. CN113645902A discloses a dental target detection method and an image integration method using dental targets, which extracts the dental focus area in the oral scan image and uses the reference point to complete the image integration of the dental target, achieving dental target detection of three-dimensional oral scan images. This technical solution has high accuracy in molar tooth target detection, but the accuracy of the target detection of the dense middle part of the teeth is significantly reduced.

[0006] Although there are currently a large number of dental feature recognition methods to assist dental diagnosis, it is still necessary to propose a small sample dental target detection method that can solve the above four problems. In view of the deficiencies of the current technology in the above four tasks, the present application proposes corresponding steps for improvement, and combines the improved steps to propose a small sample dental target detection method-4Unet target detection method. SUMMARY

[0007] The purpose of the present application is to provide a small sample, high precision dental target detection method based on dental position information classification, which generates high precision semantic edges through image augmentation, and can accurately separate single tooth contours from dental semantic edge images.

[0008] The technical scheme adopted by the present application (4Unet dental target detection method) to solve its technical problems is: (1) using targeted augmentation conforming to biological characteristics for training images to improve the performance of the dental segmentation model; (2) processing the dental semantic edge image, the dental segmentation image and the semantic segmentation image of the overlapping part between the teeth to obtain a high-precision dental semantic edge image; (3) proposing a "limited connected region" method to obtain a single tooth contour image from the dental semantic edge image through polar coordinates; (4) adding a position information-based method to classify the single tooth contour image and output the classification result of the tooth.

[0009] The specific steps are as follows:

[0010] Step (1), augmenting small sample dental images into large samples, including local area augmentation for semantic segmentation model training set and panoramic image augmentation for dental classification model.

[0011] Specifically, the degree of elastic deformation, rotation and scaling during tooth augmentation is controlled to make the tooth augmentation conform to the differences between different people's teeth in biology. The image transformation of each region in the augmentation, i.e. the transformation of the local tooth shape, "simulates" the differences between each person's teeth.

[0012] Translation and rotation can reduce the risk of overfitting and increase the robustness to some small perturbations of the input image. The image rotation process is as follows: if a pixel A in the image is to be rotated, it will become A' after rotation by θ, which requires multiplying the matrix T by the two-dimensional vector A to obtain the rotated vector A':

[0013] A→A′;

[0014]

[0015]

[0016] Elastic deformation is the core of sample expansion. According to the most common deformation in human tissue, the corresponding elastic deformation degree is used, and the process is as follows: first, create a random displacement field to deform the image, that is, Δx(x, y) = rand(-1, +1), Δy(x, y) = rand(-1, +1), wherein rand(-1, +1) is a random number uniformly distributed between (-1, 1), then convolve Δx and Δy with a Gaussian function with a standard deviation of σ, wherein the process of Gaussian convolution is:

[0017] I σ = I * G σ ;

[0018]

[0019] The elastic deformation displacement field is obtained by the corresponding standard deviation σ, σ ∈ [3, 10]; the obtained displacement field is multiplied by the scale factor α to obtain the final elastic deformation displacement field, wherein α ∈ [10, 34], and the elastic deformation expansion in the range of standard deviation σ and scale factor α is different from the expansion image of the small sample image, and does not exceed the degree of normal deformation of the dental tissue. Finally, the elastic deformation field is applied to the original image to obtain the interpolated elastic deformation image.

[0020] Step (2), the images generated by the four U-Net models are "or" to solve the problem of inaccurate boundary of the tooth generated by the U-Net model, and a high-precision tooth edge image is generated.

[0021] Four different semantic segmentation networks are trained using the semantic segmentation data set expanded in step (1). M1 is a semantic edge image of the tooth manually labeled from the original image O1, M2 is a semantic segmentation image of the tooth manually labeled from the original image O1, and M4 is a semantic segmentation image of the tooth overlap part manually labeled from the original image O1. The original image O1 and the semantic edge image M1 are used to train the U-Net A model, the original image O1 and the semantic segmentation image M2 are used to train the U-Net B model, the tooth semantic segmentation image M3 generated by the U-Net B model is superimposed with the original image O1 to obtain a new image O2, the new image O2 and the semantic edge image M1 are used to train the U-Net C model, and the original image M1 and the tooth overlap part semantic segmentation image M4 are used to obtain the U-Net D model.

[0022] Then the X-ray original image of the tooth is passed through the U-Net A model to generate the semantic segmentation image M f of the tooth and the U-Net C model to generate the enhanced edge M P of the tooth, and then the generated semantic edge M e and the enhanced semantic edge M PBitwise OR operation is performed to obtain a new tooth edge M e1 The new tooth edge is overlapped with the tooth overlapping part semantic edge M generated by the U-Net D model j Bitwise OR operation is performed to obtain a high-precision tooth edge image M e2 The closure of the tooth edge is strengthened.

[0023] In step (3), a single tooth contour is extracted from the tooth semantic edge image, and a "limited connected region" method is proposed. First, scanning from left to right and from top to bottom, since the area of each tooth contour cannot be too small, a pixel point is selected as a key point for scanning every multiple pixels. After passing through a key point, it is checked whether the point belongs to the contour of the tooth. If it belongs to the target region, a limited connected region is generated. Starting from a point, it is extended in the four directions of up, down, left and right until it touches the contour edge of the tooth. The target frame data is obtained from the boundaries of the upper, lower, left and right sides of the target. The larger the area of the target frame is, the closer it is to completely enclosing a single tooth. Therefore, the non-maximum suppression algorithm (NMS) commonly used in the field of target detection is used to rank the area of the target frame as the primary weight, and the limited connected region target frame data is generated. Some targets with sizes far away from the normal value range are filtered out.

[0024] Through the generation of the limited connected region, the contour of a single tooth is generated through the polar coordinates of the two-dimensional coordinate system. First, take each target frame center point O as the polar point, draw a ray Ox as the polar axis, and start with 0° and scan 360° counterclockwise. The length of the line segment of the polar coordinates is the distance from the center point O to the contour edge of the tooth at each scale. The length and endpoint coordinates of each ray at these scales are obtained, and the contour of a single tooth is generated according to each endpoint. In order to obtain a more accurate single tooth contour image, the centroid of the initially obtained single tooth contour image is calculated, and a single tooth contour image is generated based on each centroid. After multiple cycles, the final single tooth contour image is generated from the high-precision edge image.

[0025] In step (4), a mechanism based on position information is added to generate a single tooth contour on the original image, mark the position of a single target, and retain the contour features of a single tooth. The manually annotated tooth contour is modified, and then the improved single tooth contour image is subjected to panoramic image augmentation in the first step. The large sample after augmentation is trained using a Resnet model to obtain a classification model. The image to be detected is input into the trained classification model, and the classification result is output.

[0026] Compared with the prior art, the present application has at least the following advantages and effects:

[0027] 1. The application adopts a more reasonable expansion method, and adopts local area expansion for semantic segmentation samples, so that multiple times of dental semantic segmentation images can be obtained before expansion by cutting local areas; panoramic area expansion is adopted for classification samples, position information and contour information are retained. By controlling the parameters of elastic deformation, rotation and scaling, the large samples expanded by small samples are effective samples within the deformation degree of human tooth tissue.

[0028] 2. The application forms high-precision dental semantic edge images by complementary advantages of four types of images generated by four types of models, namely dental semantic edge, dental semantic segmentation, dental overlapping part semantic segmentation and strengthened dental semantic edge, so that the dental semantic edge has good closure and high accuracy.

[0029] 3. The application proposes a "limited connected region" method, and extracts the contour of a single tooth by using only the semantic edge image through the polar coordinate method.

[0030] 4. The application solves the problem that the difference between the contours of different types of teeth is not large enough by using a mechanism based on position information, and achieves fast and accurate tooth classification by complementary advantages of tooth contour and position information.

[0031] The application can improve the efficiency of tooth target detection in the tooth diagnosis process, and lay a foundation for realizing automatic tooth detection. DETAILED DESCRIPTION

[0032] Figure 1 is the overall flow of the application;

[0033] Figure 2 is a local area expansion principle diagram in the embodiment of the application;

[0034] Figure 3 is a panoramic area expansion principle diagram in the embodiment of the application;

[0035] Figure 4 is a flowchart for obtaining a high-precision tooth edge image from a tooth image;

[0036] Figure 5 is a segmentation performance index diagram of different methods;

[0037] Figure 6 is a flowchart for extracting a single tooth contour from a tooth edge image;

[0038] Figure 7 is a schematic diagram of extracting a single tooth target frame from a dental semantic edge image in the embodiment;

[0039] Figure 8 is a schematic diagram of the polar coordinate;

[0040] Figure 9 Generate a schematic diagram on the original image for a single tooth contour image;

[0041] Figure 10 The effect diagram for modifying the manually annotated single tooth edge in the embodiment;

[0042] Figure 11 The classification result output after inputting the test image in the embodiment. DETAILED DESCRIPTION

[0043] The present application will be described in detail below with specific embodiments. The annotations of the present application have differences in English letter case, which cannot be regarded as the same letter. The following embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any form. It should be pointed out that, for those skilled in the art, without departing from the concept of the present application, a number of changes and improvements can be made, which are within the scope of protection of the present application.

[0044] As shown in Figure 1 A high-precision tooth target detection method based on small samples, the specific steps are as follows: step (1) is the targeted expansion of panoramic images and multi-region local images under small samples, step (2) is to generate high-precision edge images using four U-Net models, step (3) is single tooth extraction based on edges, and step (4) is classification based on tooth position information combined with tooth contour.

[0045] In this embodiment, the data set used is a tooth CT image publicly available on the network, including 100 RGB images, part of the images o1 is used for semantic segmentation, and the tooth contour is marked on these images o1, and the rest is used for testing.

[0046] In step (1), for the training set used for semantic segmentation, first, the semantic edge image m1 of the tooth, the semantic segmentation image m2 of the tooth, the semantic segmentation image m4 of the tooth overlapping part, the new image m3 obtained by superimposing the tooth semantic segmentation image m2 on the original image o1, and the image o3 (o3 number is equal to the number of manually annotated teeth of the image o1) are generated by manual annotation.

[0047] The new image m3 corresponds to the semantic edge image m1, the original image o1 corresponds to the semantic segmentation image m2, the original image o1 corresponds to the semantic edge image m1, the original image o1 corresponds to the tooth overlapping part semantic segmentation image m4, and the method of using local region expansion to obtain large sample local region image.

[0048] As shown in Figure 2As shown, the local region expansion specifically refers to cutting 3 images of local regions of teeth from the rectangular tooth grayscale image, and the feature difference of the images is obvious. In the expansion process, in addition to the traditional illumination change, mirror image, noise and other processing, the elastic deformation control deformation strength scale factor a e [10, 34], the elastic coefficient s e [3, 10], the maximum left rotation angle is 8, the maximum right rotation angle is 8, the scaling ratio is less than 0.2 times, and scaling, rotation, elastic deformation probability random combination are performed. All the small sample images segmented by the semantic segmentation are expanded to obtain the semantic segmentation dataset U.

[0049] Since the classification model needs position information during training, the cutting of the panoramic tooth image is skipped, as shown in Figure 3 As shown, the panoramic image expansion specifically refers to directly performing elastic deformation, rotation and scaling on the panoramic tooth image. In the expansion process, the elastic deformation control deformation strength scale factor a e [10, 34], the elastic coefficient s e [3, 10]. The maximum left rotation angle is 8, the maximum right rotation angle is 8, the scaling ratio is less than 0.2 times, and scaling, rotation, elastic deformation probability random combination are performed. All the small sample images used for classification are expanded to obtain the classification dataset O3.

[0050] In step (2), the local region expansion semantic segmentation dataset U obtained in step (1) is used for training, and the U-Net model is used for training in the present application. The ratio of the training dataset to the validation dataset is set to 8:2, the training dataset and the validation dataset are randomly divided, the network training adopts the Mini-batch gradient descent method to optimize the target loss function, the learning rate is 0.0001, the momentum parameter is 0.9, the weight decay is set to 0.0000001, and the cross entropy loss function (CrossEntropyLoss) is used to train the network until the training loss function converges (the above-mentioned training parameters can be used as an example for reference, and as long as a model with certain effect is trained in the present application). M1 is a semantic edge image of the tooth of the original image O1 manually labeled, and M2 is a semantic segmentation image of the tooth of the original image O1 manually labeled. The U-Net A model is trained using the original image O1 and the semantic edge image M1, the U-Net B model is trained using the original image O1 and the semantic segmentation image M2, the tooth semantic segmentation image M3 generated by the U-Net B model, the tooth semantic segmentation image M3 channel superimposed original image O1 obtains a new image O2, the U-Net C model is trained using the new image O2 and the semantic edge image M1, and the U-Net D model is obtained using the original image M1 and the tooth overlapping part semantic segmentation image M4.

[0051] As shown in Figure 4 , the X-ray raw image of the tooth is processed by the U-Net A model to generate the semantic edge M e of the tooth, and the U-Net C model to generate the reinforced edge MP of the tooth, and then the generated semantic edge M e and the reinforced semantic edge MP are subjected to "bitwise or" processing to obtain a new tooth edge M e1 , and then the new tooth edge is subjected to "bitwise or" processing with the tooth overlapping part semantic edge generated by the U-Net D model to obtain a high-precision tooth edge image M e2 , which ensures the closure of the tooth edge. As shown in Figure 5 , the final supplementary segmentation image M e2 generated in this step is shown in

[0052] The closure rate comparison is shown in the following table:

[0053]

[0054] The semantic segmentation evaluation indicators are Mean Pixel Accuracy (MPA), Mean Intersection over Union (mlOU), and the probability of complete closure of the single tooth contour (closure rate)

[0055] As shown in Figure 6 , in step (3), the tooth edge image M e2 generated in step (2) is used to extract the single tooth contour from the tooth semantic edge image, and a "limited connected region" method is proposed. First, scanning from left to right and from top to bottom, since scanning pixel by pixel is too time-consuming, the area of each tooth contour will not be too small, so every multiple of pixels is scanned as a key point, and whether this place belongs to the contour of the tooth is checked every time a key point is passed. If it is in the tooth contour, it belongs to the target area, as shown in Figure 7 , if it is a target area, start to generate a limited connected region, starting from a point, extending in four directions up, down, left and right until it touches the edge of the tooth contour, and the target frame data is obtained from the boundaries of the upper, lower, left and right sides of the target. The larger the area of such target frame, the closer it is to completely enclosing a single tooth, so the non-maximum suppression algorithm (NMS) commonly used in the target detection field is used to rank the area of the target frame as the primary weight, and the limited connected region target frame data is generated. Then, some targets with sizes far away from the normal value range are screened out.

[0056] Figure 8For the polar coordinate overview map, through the generation of the above limited connected region, the contour of a single tooth can now be generated by the polar coordinates of the two-dimensional coordinate system, first taking each target frame center point O as the polar point, drawing a ray Ox as the polar axis, taking 0° as the starting point, and scanning 360° counterclockwise, the length of the polar coordinate line segment is the distance from the center point O to the contour edge of the tooth at each scale, and the length and endpoint coordinates of each ray at these scales are obtained, and the contour of a single tooth is generated according to each endpoint. In order to obtain a more accurate single tooth contour image, the centroid of the initially obtained single tooth contour image is calculated, and a single tooth contour image is generated based on each centroid. After three cycles, the final single tooth contour image is generated from the high-precision edge image.

[0057] As shown in Figure 9 , in step (4), the single tooth contour highlighted display on the original image exists position information, as shown in Figure 10 , the single tooth edge highlighted image o4 obtained in step (3) is compared with the single tooth edge highlighted image o3 manually labeled, and it is found that there is a small difference between the single tooth contour manually labeled and the single tooth contour generated by the algorithm. Considering this problem, the manually labeled image is modified. The modified single tooth edge highlighted image is expanded into a panoramic image according to the panoramic image expansion method of the first step, and a large sample data image O5 is obtained by expanding a certain multiple. Finally, the large sample image O5 is trained using the Resnet model, the ratio of the training data set to the test data set is set to 8:2, the small batch (Mini-batch) gradient descent method is used to optimize the target loss function, the learning rate (learning rate) is set to 0.1, gamma = 0.1, epoch is taken in the interval [200, 500], and every time the validation set loss does not decrease after 10 epochs of training, the learning rate becomes 0.1 of the original, the momentum parameter (momentum) is taken as 0.9, the target loss function in this example is the cross entropy loss function (CrossEntropyLoss), and the classification model Re34et a is obtained after all epochs are iterated (the above trained parameters can be used as an example for reference, and as long as a model with a certain effect is trained in this application).

[0058] As shown in Figure 11 , the tooth image used for testing is input, the image is generated into a high-precision tooth edge by the four U-Net networks in step (2), and then the single tooth contour highlighted image o4 is generated by step (3), and these images are classified by the classification model Re34et a in step (4), and the category of each tooth is classified.

[0059] In order to reduce the time spent in tooth target detection in diagnosis, reduce the labor cost and ensure the accuracy, four steps are combined to ensure higher efficiency, and in actual use, the steps can be adjusted according to specific conditions to obtain higher accuracy.

[0060] The specific embodiments of the present application are described above. The present application is not limited to the specific embodiments described above, and various changes or modifications can be made by those skilled in the art within the scope of the claims, which does not affect the essential content of the present application. The embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily without conflict.

Claims

1. A small sample-based high-precision tooth target detection method, characterized in that: Specifically: (1) The training image is amplified in accordance with the biological characteristics, so as to improve the performance of the tooth segmentation model; (2) The tooth semantic edge image, the tooth segmentation image and the semantic segmentation image of the overlapping part between the teeth are processed to obtain a high-precision tooth semantic edge image; (3) A "limited connected region" method is proposed, and the single tooth contour image is obtained from the tooth semantic edge image through polar coordinates; (4) A method based on position information is added to classify the single tooth contour image and output the classification result of the tooth; The step (3) is specifically: extracting a single tooth contour from the tooth semantic edge image, and proposing a "limited connected region" method; first, scanning from left to right and from top to bottom, and selecting a pixel point as a key point every multiple pixels; whether the key point is in the tooth contour is checked, and if it is in the tooth contour, it is in the target region; if it is in the target region, a limited connected region is generated, which starts from a point and extends to the upper, lower, left and right directions until it touches the edge of the tooth contour; the target box data of the limited connected region is obtained from the boundaries of the upper, lower, left and right sides of the target, and the non-maximum suppression algorithm commonly used in the field of target detection is used to rank the area of the target box as the primary weight, and the limited connected region target box data is generated to filter out the targets whose size deviates from the normal value range; According to the generation of the limited connected region, the contour of a single tooth is generated through polar coordinates of a two-dimensional coordinate system; first, taking each target box center point O as a polar point, drawing a ray Ox as a polar axis, and starting from 0°, the contour of a single tooth is generated by scanning 360° counterclockwise; the length of each scale is the distance from the center point O to the contour edge of the tooth, and the length and endpoint coordinates of each ray are obtained at each scale; the contour of a single tooth is generated according to each endpoint; in order to obtain a more accurate single tooth contour image, the centroid of the initially obtained single tooth contour image is calculated, and a single tooth contour image is generated based on each centroid; after multiple cycles, the final single tooth contour image is generated from the high-precision edge image.

2. The small sample based high-precision tooth target detection method of claim 1, wherein: The step (1) is specifically: expanding the small sample tooth image into a large sample, including local region expansion for the semantic segmentation model training set and panoramic image expansion for the tooth classification model; by controlling the degree of elastic deformation, rotation and scaling during tooth expansion, the tooth expansion conforms to the biological differences between different people's teeth.

3. The small sample based high precision tooth target detection method of claim 1, wherein: The step (2) is specifically: the images generated by the four U-Net models are subjected to "bitwise OR" to solve the problem of inaccurate tooth boundary generated by the U-Net model, and a high-precision tooth edge image is generated; Four different U-Net models were trained using the semantic segmentation dataset after local region augmentation in step (1). Original images with manual annotation The semantic edge image of teeth, Original images with manual annotation Semantic segmentation image of teeth, Original images with manual annotation Semantic segmentation image of overlapping teeth; using the original image. With semantic edge images The U-Net A model was trained using the original images. With semantic segmentation images The U-Net B model was trained, and the tooth semantic segmentation images generated by the U-Net B model were used. Channel overlay of original image Get a new image New image With semantic edge images The U-Net C model was trained using the original images. Semantic segmentation image of overlapping parts with teeth Obtain the U-Net D model; Then the X-ray raw image of the tooth is processed by the U-Net A model to generate a semantic segmentation image of the tooth and the U-Net C model to generate a reinforced edge of the tooth Then the generated semantic edge and the reinforced semantic edge are subjected to "bitwise or" processing to obtain a new tooth edge Then the new tooth edge is subjected to "bitwise or" processing with the tooth overlapping part semantic edge generated by the U-Net D model to obtain a high-precision tooth edge image , which enhances the closure of the tooth edge.

4. The small sample based high precision tooth target detection method of claim 1, wherein: The step (4) is specifically: adding a position information-based mechanism, generating the contour of a single tooth on the original image, marking the position of the single target and retaining the contour features of the single tooth; modifying the manually-annotated tooth contour, then performing panoramic image augmentation on the improved single tooth contour image in the step (1), and training the augmented large sample using a Resnet model to obtain the classification model of the application; inputting the picture to be detected into the trained classification model, and outputting a classification result.

Images