3D printing false tooth model change detection method

Abstract

The invention relates to the technical field of image processing change detection, and particularly discloses a 3D printing denture model change detection method which comprises the following steps: S1, shooting a full-page printing top view after 3D printing processing, taking the full-page printing top view as an input A, intercepting a typesetting picture in 3D printing typesetting software, and taking the typesetting picture as an input B; s2, performing gray processing, Gaussian filtering denoising, adaptive canny edge detection and contour area screening on the input A to obtain a maximum outer contour, namely the outer contour of the 3D printing chassis, calculating a minimum envelope rectangle of the contour, performing transmission transformation on the rectangle and a square to obtain a corrected 3D printing chassis A1, and performing identical processing on the input B to obtain B1; the 3D printing denture model change detection method has the advantages of high image registration and change detection precision, insensitivity to light and shadow changes, high detection speed and low requirement on configuration of photographic equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing change detection, and particularly to a method for detecting changes in 3D printed denture models. Background Art

[0002] With the development of digital technology, 3D printing has been widely used in the manufacturing industry. It has the advantages of high precision, high efficiency, customization, and reduction of material loss, and has become the mainstream and trend of manufacturing technologies in many fields (such as denture manufacturing technology). Currently, most factory production adopts manual supervision, which lacks objectivity, effectiveness, and timeliness, and there are often behaviors of employees taking on private jobs. Change detection can provide an effective means of production supervision. The specific manifestation of production anomalies is that the processed parts do not match the typesetting pictures, such as the increase or decrease in quantity or replacement. By inputting the processed pictures and typesetting pictures, the changed parts are detected, and a binary map of change prediction is output.

[0003] Change detection is usually used in the field of remote sensing. By using multi-temporal remote sensing images of the same geographical location, the detection and analysis of surface soil cover types, urban street changes, building complex changes, etc. are carried out. The targets in its multi-temporal images have the same angular, distance, and radiation changes in perspective, but for each target in the multi-temporal images, its physical self does not change. In the dual-input images processed by 3D printing, the parts in the typesetting picture (input B) are standard top views, which are plane projections and computer-generated pictures, and the parts do not have supports. While the physical photo of the 3D printed processed object (input A) is taken by a mobile phone or other photographic equipment, which is a central projection and a physical picture, and the parts are lifted by the printed supports. In the top view, each part has warping deformations at different angles and directions, which causes different changes in the corresponding parts between A and B in terms of physical meaning, and different changes also occur between different feature point pairs, bringing many difficulties to image registration and change detection, such as the failure of feature point matching and the decrease in prediction accuracy caused by the warping deformation of parts. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for detecting changes in 3D printed denture models, which can complete image registration in the case of feature point matching failure and complete change detection in the case of part warping deformation.

[0005] To achieve the above purpose, the present invention provides a method for detecting changes in 3D printed denture models, including the following steps:

[0006] S1. Take a top view of the whole plate of the 3D printed processed denture, use this top view of the whole plate as input A, intercept the typesetting picture of the denture in the 3D printing typesetting software, and use this typesetting picture as input B;

[0007] S2. Perform grayscale processing and Gaussian filtering denoising on input A, and use adaptive Canny edge detection and contour area screening to process input A to obtain the maximum outer contour, that is, the outer contour of the 3D printing chassis. Calculate the minimum bounding rectangle of the obtained maximum outer contour, and perform a projective transformation on the minimum bounding rectangle and a square to obtain the corrected 3D printing chassis A 1 , perform the same processing on input B to obtain B 1 ;

[0008] S3. Set the image sizes of A 1 and B 1 to 1024×1024, and convert them into binary images. Calculate the cosine similarity of the current orientations of A 1 and B 1 . Rotate A 1 1 - 3 times by 90°. Calculate the cosine similarity of the current orientations of A 1 and B 1 each time it is rotated. Align the orientations with the maximum similarity, and complete the image registration of A 1 and B 1 to obtain the registered image pair A 2 , B 2 ;

[0009] S4. Collect and produce multiple datasets similar to A 2 , B 2 and with varying parts, and label the varying parts; Build a deep learning change detection network model based on CBAM - UNet, use the above datasets for training, and save the model; Use the model to predict the change detection of dual - input images similar to A 2 , B 2 and output the predicted binary image.

[0010] Preferably, in S2, obtain the four vertices of the minimum bounding rectangle of the maximum outer contour of the input image, calculate the homography matrix of these four vertices to the four vertices of the square, and perform a projective transformation on the image within the range of this homography matrix.

[0011] Preferably, in S3, let the binary image of A 1 be vector a, the binary image of B 1 be vector b, and the formula for calculating the cosine similarity of the current orientations of A 1 and B 1 is as follows:

[0012]

[0013] where, · is the dot - product calculation, and × is the cross - product calculation.

[0014] Preferably, in S3, image registration specifically refers to the double-input images after the printing chassis coincides, calculating the cosine similarity between the current orientations of the two, and rotating A 1 three times by 90°, calculating the cosine similarity between the current orientation and B 1 each time it rotates, obtaining the cosine similarities of 4 orientations and the orientation with the maximum cosine similarity, and rotating A 1 to the orientation with the maximum cosine similarity to complete the alignment with the orientation of B 1 .

[0015] Preferably, in S4, a deep learning change detection network model based on CBAM-UNet is built. Taking UNet as the architecture, a CBAM attention module composed of a channel attention module and a spatial attention module is added during the downsampling process of UNet; let F be the input feature, F ∈ R C*H*W , the one-dimensional convolution M C ∈ R C*1*1 of the channel attention module, the two-dimensional convolution M S ∈ R 1*H*W of the spatial attention module. The feature map obtained through the channel attention module is F’, and the final output is F”. The feature map calculation formula is as follows:

[0016] M C (F) = σ(MLP(AvgPool(F)) + MLP(MaxPool(F)));

[0017] M S (F) = σ(f 7x7 ([AvgPool(F); MaxPool(F)]));

[0018]

[0019] where M C (·) is a one-dimensional convolution operation, M S (·) is a two-dimensional convolution operation, AvgPool(·) is an average pooling operation, MaxPool(·) is a max pooling operation, MLP(·) is a multi-layer perceptron operation, in R C*H*W , R is a matrix, C is the number of channels, H and W are the height and width of the image respectively, σ is the sigmoid activation function, f is a convolution, and 7x7 is the convolution box size.

[0020] Preferably, in S4, after inputting the dataset into the network model, the network model is trained using the dataset. The cross-entropy loss function is used to calculate the difference between the change detection result and the change detection label, and the backpropagation algorithm is used to update the model parameters. The current model parameters are used to evaluate the validation set and assign a score. After multiple rounds of learning, the model parameters with the highest score are selected and saved as the optimal model, and the optimal model is used for prediction.

[0021] Therefore, the present invention adopts the above-mentioned method for detecting changes in a 3D printed denture model, and the beneficial effects are as follows:

[0022] (1) The present invention first applies change detection in the field of 3D printing, overcoming the problems of feature point matching failure in the image registration process caused by the particularity of 3D printing (the warping deformation of parts in the top view caused by 3D printing supports) and the misjudgment of the changed part in change detection.

[0023] (2) The present invention provides an effective means of production supervision, which has high image registration and change detection accuracy, is insensitive to light and shadow changes, has a fast detection speed, and has low requirements for the configuration of photographic equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 is the overall flowchart of an embodiment of the method for detecting changes in a 3D printed denture model according to the present invention;

[0025] Figure 2 is the dual-input picture of an embodiment of the method for detecting changes in a 3D printed denture model according to the present invention;

[0026] Figure 3 is the picture A after processing the dual input of an embodiment of the method for detecting changes in a 3D printed denture model according to the present invention 1 , B 1 ;

[0027] Figure 4 is the picture A after registration of the dual input of an embodiment of the method for detecting changes in a 3D printed denture model according to the present invention 2 , B 2 ;

[0028] Figure 5 is the structure diagram of the CBAM-UNet change detection network of an embodiment of the method for detecting changes in a 3D printed denture model according to the present invention;

[0029] Figure 6 is the structure of the CBAM module of an embodiment of the method for detecting changes in a 3D printed denture model according to the present invention;

[0030] Figure 7It is the output predicted change binary map of an embodiment of the 3D printed denture model change detection method of the present invention. Detailed implementation manners

[0031] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0032] Unless otherwise defined, the technical terms or scientific terms used in the present invention shall have the ordinary meanings understood by those of ordinary skill in the field to which the present invention belongs.

[0033] Embodiment

[0034] As Figure 1 shown, a 3D printed denture model change detection method includes the following steps:

[0035] S1. As Figure 2 shown, take the top view of the whole plate after 3D printing and processing, use this top view as input A, intercept the layout picture in the 3D printing layout software, and use this picture as input B;

[0036] S2. As Figure 3 shown, perform grayscale processing on input A, denoise it with Gaussian filtering, use adaptive canny edge detection and contour area screening to obtain the largest outer contour, that is, the outer contour of the 3D printed chassis (a square-like shape with 4 rounded corners), calculate the minimum bounding rectangle of this contour, perform a projective transformation on this rectangle and the square to obtain the corrected 3D printed chassis A 1 , perform the same processing on input B to obtain B 1 ;

[0037] S3. As Figure 4 shown, set the image sizes of A 1 and B 1 to 1024×1024, and convert them into binary maps. Calculate the cosine similarity of the current orientations of A 1 and B 1 . Rotate A 1 3 times by 90°, calculate the cosine similarity each time, store it in a list, obtain the orientation combination with the largest cosine similarity, rotate A 1 to this orientation, and complete the image registration of A 1 and B 1 to obtain the registered image pair A 2 , B 2 ;

[0038] S4. As Figure 5-6 shown, collect and make multiple groups in the form of A 2 , B 2And a data set with variable parts, and label the variable parts. Build a deep learning change detection network based on CBAM-UNet, with UNet as the architecture, and add a CBAM attention module composed of a channel attention module and a spatial attention module during the downsampling process of UNet; Let F be the input feature, F ∈ R C*H*W , the one-dimensional convolution M of the channel attention module C ∈ R C*1*1 , the two-dimensional convolution M of the spatial attention module S ∈ R 1*H*W , the feature map passing through the channel attention module is F’, and the final output is F”, and the feature map calculation formula is as follows:

[0039] M C (F) = σ(MLP(AvgPool(F)) + MLP(MaxPool(F)));

[0040] M S (F) = σ(f 7x7 ([AvgPool(F); MaxPool(F)]));

[0041]

[0042]

[0043] Among them, M C (·) is a one-dimensional convolution operation, M S (·) is a two-dimensional convolution operation, AvgPool(·) is an average pooling operation, MaxPool(·) is a maximum pooling operation, MLP(·) is a multi-layer perception operation, R C*H*W In, C is the number of channels, and H and W are the height and width of the image respectively.

[0044] Use the data set for training and save the model. Use the model to predict the change detection of double-input images in the form of A 2 , B 2 , output the predicted binary map, remove the noise of the small-area pseudo-change region in the predicted map, and the final output result is as Figure 7 shown, as Figure 7 can be seen, the prediction result has high accuracy.

[0045] Therefore, the present invention adopts the above-mentioned method for detecting changes in a 3D printed denture model, which overcomes the failure of feature point matching in the image registration process caused by the particularity of 3D printing (the warping deformation of parts in the top view caused by 3D printing supports) and the judgment of it as a variable part in change detection, etc.; it is not sensitive to light and shadow changes, has a fast detection speed, and has low configuration requirements for the photographic equipment.

[0046] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions of the present invention or make equivalent replacements, and these modifications or equivalent replacements do not enable the modified technical solutions to depart from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for detecting changes in a 3D printed denture model, characterized in that: The following steps are involved: S1, photographing a top view of the entire printed denture after 3D printing, using the top view as input A, intercepting a layout image of the denture in the 3D printing layout software, using the layout image as input B; S2, grayscale processing and Gaussian filtering are performed on input A, and input A is processed using adaptive canny edge detection and contour area screening to obtain the maximum outer contour, that is, the outer contour of the 3D printed chassis, and the minimum envelope rectangle of the maximum outer contour is calculated. The minimum envelope rectangle is subjected to a transmission transformation with the square to obtain the corrected 3D printed chassis A1, and the same processing is performed on input B to obtain B1; S3, set the image size of A1 and B1 to 1024×1024, and convert them into binary images, calculate the cosine similarity of the images facing A1 and B1, rotate A1 1-3 times by 90°, calculate the cosine similarity of the images facing A1 and B1 each time, align the orientation with the maximum similarity, complete the image registration of A1 and B1, and obtain the registered image pair A2 and B2; S4. Collect and create multiple sets of data sets that are similar to A2 and B2 and have changed parts, and label the changed parts; build a deep learning change detection network model based on CBAM-UNet, use the above data sets for training, and save the model; use the model to predict change detection on dual input images similar to A2 and B2, and output the predicted binary image.

2. A 3D printed denture model change detection method according to claim 1, characterized in that: In S2, the four vertices of the minimum enveloping rectangle of the maximum outer contour of the input image are obtained, the homography matrix from the four vertices to the four vertices of the square is calculated, and the image within the range of the homography matrix is ​​subjected to a transmission transformation.

3. A 3D printed denture model change detection method according to claim 1, characterized in that: In S3, let the binary image of A1 be vector a, and the binary image of B1 be vector b. The formula for calculating the cosine similarity of the images facing A1 and B1 is as follows: Among them, · is the dot product calculation, and × is the cross product calculation.

4. A 3D printed denture model change detection method according to claim 3, characterized in that: In S3, image registration is to print the double input images after the chassis overlap, calculate the cosine similarity of the current orientation of the two, rotate A1 three times by 90 degrees ° , calculate the cosine similarity between the orientation and B1 each time it is rotated, obtain the cosine similarities of the four orientations and the orientation with the maximum cosine similarity, rotate A1 to the orientation with the maximum cosine similarity, and complete the orientation alignment with B1.

5. A 3D printed denture model change detection method according to claim 1, characterized in that: In S4, a deep learning change detection network model based on CBAM-UNet is built. UNet is used as the architecture, and a CBAM attention module composed of a channel attention module and a spatial attention module is added to the downsampling process of UNet. Let F be the input feature, F∈R C *H*W , channel attention module one-dimensional convolution M C ∈R C*1*1 , spatial attention module 2D convolution M S ∈R 1*H*W , the feature map passing through the channel attention module is F', and the final output is F", the feature map calculation formula is as follows: M C (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F))); M S (F)=σ(f 7x7 ([AvgPool(F);MaxPool(F)])); Among them, M C (·) is a one-dimensional convolution operation, M S (·) is a two-dimensional convolution operation, AvgPool(·) is an average pooling operation, MaxPool(·) is a maximum pooling operation, MLP(·) is a multi-layer perception operation, R C*H*W In the figure, R is a matrix, C is the number of channels, H and W are the height and width of the image, σ is the sigmoid activation function, f is convolution, and 7x7 is the convolution box size.

6. A 3D printed denture model change detection method according to claim 1, characterized in that: In S4, after the dataset is input into the network model, the network model is trained using the dataset, the cross entropy loss function is used to calculate the difference between the change detection results and the change detection labels, and the back propagation algorithm is used to update the model parameters. The validation set is evaluated and scored using the current model parameters. After multiple rounds of learning, the model parameters with the highest score are selected and saved as the optimal model, and the optimal model is used for prediction.

Images