A three-dimensional image reconstruction method and system based on dictionary learning

Abstract

The application discloses a three-dimensional image reconstruction method and system based on dictionary learning, and relates to the technical field of image processing.The method comprises the following steps: obtaining original ultrasonic field data of a rotating scanning workpiece to be measured; obtaining a low-resolution image through filtering, frequency domain transformation, secondary filtering by a ramp function and inverse transformation; constructing an external dictionary and an internal dictionary based on an open-source high-resolution CT data set and the low-resolution image respectively; coupling dictionary features in the time domain and the frequency domain by using an optimization function, iteratively optimizing a mapping relationship by using a FISTA algorithm, and reconstructing a tomographic image based on compression sensing; and finally, splicing high-resolution three-dimensional images in height order.The application realizes the super-resolution reconstruction of ultrasonic images under the double constraints of the frequency domain and the time domain by combining dictionary learning and compression sensing, and significantly improves the three-dimensional imaging quality.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a three-dimensional image reconstruction method and system based on dictionary learning. Background Technology

[0002] Ultrasound imaging, also known as ultrasound imaging or ultrasound scanning, is an imaging technique that uses high-frequency sound waves to create echoes within an object being scanned. Ultrasound imaging equipment generates high-frequency sound waves, which, when they enter the object and encounter different tissues and organs, are reflected back, forming echoes. These echoes are received by the equipment and converted into images. Different tissues reflect sound waves differently, thus appearing as different brightness levels on the screen, resulting in a clear image of the internal structure. The main advantage of ultrasound imaging is its safety. Because it does not use radiation, it is safe for the user. The portability and relatively low cost of ultrasound imaging also make it a preferred diagnostic tool in many medical settings.

[0003] Ultrasonic computed tomography (CT) belongs to the field of ultrasound imaging. It typically consists of an ultrasound array composed of ultrasound probes, which uses specific algorithms to ensure the regular emission and reception of sound waves. The task of ultrasound computed tomography is to reconstruct an image from the received reflected and transmitted signals. One method uses ultrasound reflected signals for imaging, called the echo-reflection method; the other method mainly uses transmitted signals for imaging, called the full waveform inversion method.

[0004] Ultrasound computed tomography (CT) is generally used in the field of medical imaging. This technology is used in obstetrics and gynecology to monitor fetal development and in cardiology to assess cardiac structure and function. Furthermore, ultrasound computed tomography is widely used in the examination of organs such as the abdomen, liver, breast, and thyroid gland, helping doctors detect tumors, cysts, and other lesions. Compared to X-rays and CT scans, ultrasound computed tomography is radiation-free, highly safe, and suitable for patients requiring frequent examinations. Beyond medical imaging, there is also a pressing need in other industries for the detection of internal defects in industrial fasteners using ultrasound scanning devices. Therefore, there is an urgent need for a method to transfer imaging techniques from the medical field to industrial applications. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a three-dimensional image reconstruction method and system based on dictionary learning. This method uses ultrasonic tomography to perform three-dimensional ultrasonic imaging on the workpiece under test to identify internal defects.

[0006] The specific plan is as follows:

[0007] On the one hand, a dictionary-based 3D image reconstruction method includes:

[0008] S1, acquire the original ultrasonic field data generated by the workpiece under test during the rotational scanning process;

[0009] S2, the original ultrasonic field data is arranged according to the angular direction and the first filtering is performed to obtain the first signal matrix. Based on the first signal matrix, the original ultrasonic field data is transformed in the frequency domain to obtain the frequency domain data of the original ultrasonic field data. The frequency domain data is filtered twice using the ramp function to obtain the second signal matrix. The second signal matrix is ​​stitched together and inversely transformed to obtain a low-resolution image.

[0010] S3: Obtain each slice image from the open-source high-resolution CT dataset NIST and perform preset size adjustment and equal-interval segmentation to obtain feature blocks that constitute the minimum elements of the outer dictionary. Cluster the feature blocks that constitute the minimum elements of the outer dictionary to obtain several feature groups, which constitute the outer dictionary. Perform preset size adjustment and equal-interval segmentation on the obtained low-resolution image to obtain feature blocks that constitute the minimum elements of the inner dictionary. Use the feature blocks that constitute the minimum elements of the inner dictionary as dictionary elements to obtain the inner dictionary.

[0011] S4. Based on the internal and external dictionaries, an optimization function is used to couple each feature block in the internal and external dictionaries with the low-resolution image at both the time and frequency domain levels to obtain coupled dictionary features. The mapping relationship of the coupled dictionary features is optimized using an optimization function that correlates the time and frequency domains. Based on the optimized mapping relationship and the coupling loss function, the FISTA algorithm is used to iterate the coupled dictionary features to obtain the optimal mapping relationship between dictionary elements and the low-resolution image. When the coupling loss function reaches its minimum through iteration, the obtained optimal mapping relationship is used to reconstruct the low-resolution image based on compressed sensing to obtain the reconstructed tomographic image.

[0012] S5 stitches the reconstructed tomographic images in height order to obtain a high-resolution three-dimensional image of the workpiece under test.

[0013] Furthermore, the process of acquiring the original ultrasonic field data specifically includes:

[0014] S11, Pour liquid into the test tank to cover the ultrasonic array, align the geometric center of the workpiece under test with the rotation center of the linear module, and make the geometric center of the workpiece under test, the rotation center of the linear module, and the center of the motor shaft on the same axis.

[0015] S12, the transmitting array receives the transmission sequence signal and wave function signal transmitted by the control center, the receiving array receives the receiving sequence signal transmitted by the control center, and the motor driver receives the motor pulse ratio signal transmitted by the control center. Based on the transmission sequence signal, wave function signal, receiving sequence signal and motor pulse ratio signal, the drive motor rotates by one step angle and a set of ultrasonic signals penetrates the workpiece under test and is absorbed by the receiving array, thereby obtaining the original ultrasonic field data.

[0016] Furthermore, for each step angle rotated by the drive motor, a set of ultrasonic signals penetrates the workpiece under test and is absorbed by the receiving array, thereby obtaining the original ultrasonic field data. The calculation formula is as follows:

[0017] ;

[0018] in, This represents the obtained raw ultrasonic field data; Indicates the relative angle of the scanned data; This represents the signal amplitude after projection transformation; y represents the position in the original spatial plane; A function representing the geometric shape of the object being measured.

[0019] Furthermore, methods for arranging the raw ultrasonic field data by angular direction and then filtering include: mean filtering, first-order filtering, or Gaussian filtering.

[0020] Furthermore, the formula for calculating low-resolution images is as follows:

[0021] ;

[0022] ;

[0023] ;

[0024] in, This represents the obtained raw ultrasonic field data; Represents the frequency domain transform function; Indicates the window function; Indicates angular frequency; and These represent the components of the angular frequency on the horizontal axis and the components on the vertical axis, respectively. represents an imaginary number; This represents the frequency domain function after ramp filtering; Represents the real number field.

[0025] Furthermore, the construction methods for external and internal dictionaries are as follows:

[0026] For each slice of data in the open-source high-resolution CT dataset The image size was adjusted to 256*256 pixels. Each slice was divided into 1024 8*8 pixel feature blocks. These feature blocks were then clustered using the KNN algorithm to obtain j feature groups, with an external dictionary. Represented as ,in This represents the feature block that constitutes the smallest element of the external dictionary. Represents several feature groups after clustering;

[0027] The internal dictionary V generated from the low-resolution image contains only one data slice. Data slicing After resizing the image to 256*256, it was segmented at equal intervals, resulting in 1024 8*8 feature blocks. The internal dictionary... Represented as , This represents the feature block that constitutes the smallest element of the internal dictionary.

[0028] Furthermore, optimization functions are used to couple each feature block in the internal and external dictionaries with the low-resolution image at both the time and frequency domain levels to obtain the coupled dictionary features. The calculation formula is as follows:

[0029]

[0030]

[0031] in, Dictionary features indicating coupling; , and Indicates hyperparameters; Indicates an external dictionary; Indicates an internal dictionary; and This indicates that a frequency domain transformation is performed on the feature groups in the dictionary; This represents the frequency domain transform function.

[0032] Furthermore, the calculation formula for the FISTA algorithm is as follows:

[0033] ;

[0034] ;

[0035] ;

[0036] in, This represents the intermediate value of the mapping relationship after the k-th round of optimization; Indicates the step size adjustment factor; This represents the final value of the mapping relationship after the k-th round of optimization; Indicates hyperparameters; ( ) indicates that the coupled dictionary features are transposed; Multiplying the dictionary matrix by the mapping relationship of the (k-1)th round, we get the intermediate state image matrix of the (k-1)th round. This indicates a low-resolution image.

[0037] Furthermore, the reconstructed tomographic images are stitched together in height order, specifically as follows:

[0038] The current position of the workpiece is marked as the original position; the unit distance of each workpiece translation is recorded, and the imaging data before and after a single translation is compared by a time-series algorithm to check whether the reconstructed image after translation includes the internal information of the workpiece; if it is still within the imaging range of the workpiece, the reconstructed image is stitched together at a unit distance in the third axis direction; otherwise, the stitching is terminated to obtain the final reconstructed tomographic image.

[0039] On the other hand, a dictionary-based 3D image reconstruction system includes:

[0040] The ultrasonic field data acquisition module is used to acquire the raw ultrasonic field data generated by the workpiece under test during the rotational scanning process;

[0041] The low-resolution image acquisition module is used to arrange the original ultrasonic field data according to the angular direction and then filter it to obtain the filtered signal matrix. Based on the signal matrix, the original ultrasonic field data is transformed in the frequency domain to obtain the frequency domain data of the original ultrasonic field data. The frequency domain data is filtered twice using a ramp function to obtain the signal matrix after secondary filtering. The signal matrix after secondary filtering is stitched together and inversely transformed to obtain the low-resolution image.

[0042] The internal and external dictionary generation modules are used to acquire and perform preset size adjustment and equal-interval segmentation on each slice data image in the open-source high-resolution CT dataset NIST to obtain feature blocks that constitute the minimum elements of the external dictionary. The feature blocks that constitute the minimum elements of the external dictionary are clustered to obtain several feature groups, which constitute the external dictionary. The modules also perform preset size adjustment and equal-interval segmentation on the obtained low-resolution images to obtain feature blocks that constitute the minimum elements of the internal dictionary. The feature blocks that constitute the minimum elements of the internal dictionary are directly used as dictionary elements to obtain the internal dictionary.

[0043] The tomographic image reconstruction module uses an optimization function to couple each feature block in the internal and external dictionaries with the low-resolution image at both the time and frequency domain levels, based on the constructed internal and external dictionaries, to obtain coupled dictionary features. The mapping relationship of the coupled dictionary features is then optimized. Based on the optimized mapping relationship and the coupling loss function, the FISTA algorithm is used to iterate the coupled dictionary features to obtain the optimal mapping relationship between dictionary elements and the low-resolution image. When the coupling loss function reaches its minimum through iteration, the obtained optimal mapping relationship is used to reconstruct the low-resolution image based on compressed sensing, resulting in the reconstructed tomographic image.

[0044] The high-resolution image acquisition module is used to stitch together the reconstructed tomographic images in height order to obtain a high-resolution three-dimensional image of the workpiece under test.

[0045] The present invention adopts the above technical solution and has the following beneficial effects:

[0046] (1) This invention constructs an external dictionary and an internal dictionary, and uses an optimization function to perform dual coupling of the two types of dictionaries at the time and frequency domain levels to obtain the coupled dictionary features. On this basis, the mapping relationship is iteratively optimized through the FISTA algorithm, and low-resolution images are reconstructed based on compressed sensing, which can effectively restore high-frequency detail information in the image and significantly improve the resolution and imaging quality of the reconstructed image;

[0047] (2) In this invention, the high-frequency detail information of the external dictionary and the low-frequency structural information of the internal dictionary are effectively integrated through dual coupling optimization of the time domain and frequency domain, making the learning of the mapping relationship more accurate, the reconstruction process more adaptable to different types of workpieces, and effectively improving the accuracy and robustness of the reconstruction algorithm.

[0048] (3) In the process of optimizing the mapping relationship, the present invention uses FISTA for iterative solution. After the coupling loss function reaches the minimum, the optimal mapping relationship is obtained. Then, image reconstruction is completed based on compressed sensing, which significantly reduces the number of iterations and computational overhead. While ensuring the reconstruction quality, it improves the image reconstruction efficiency and is suitable for the real-time requirements in actual industrial inspection scenarios. Attached Figure Description

[0049] Figure 1 This is a flowchart of the dictionary-based 3D image reconstruction method according to an embodiment of the present invention;

[0050] Figure 2 This is a schematic diagram of the dictionary mapping relationship matrix in an embodiment of the present invention;

[0051] Figure 3 This is a diagram of a dictionary-based 3D image reconstruction system according to an embodiment of the present invention. Detailed Implementation

[0052] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0053] like Figure 1 As shown, the present invention provides a 3D image reconstruction method based on dictionary learning, comprising:

[0054] S1, acquire the raw ultrasonic field data generated by the workpiece under test during the rotational scanning process.

[0055] Specifically, the process of acquiring raw ultrasonic field data includes:

[0056] S11, Pour liquid into the test tank to cover the ultrasonic array, align the geometric center of the workpiece under test with the rotation center of the linear module, and make the geometric center of the workpiece under test, the rotation center of the linear module, and the center of the motor shaft on the same axis.

[0057] S12, the transmitting array receives the transmission sequence signal and wave function signal transmitted by the control center, the receiving array receives the receiving sequence signal transmitted by the control center, and the motor driver receives the motor pulse ratio signal transmitted by the control center. Based on the transmission sequence signal, wave function signal, receiving sequence signal and motor pulse ratio signal, the drive motor rotates by one step angle and a set of ultrasonic signals penetrates the workpiece under test and is absorbed by the receiving array, thereby obtaining the original ultrasonic field data.

[0058] Specifically, for each step angle rotated by the drive motor, a set of ultrasonic signals penetrates the workpiece under test and is absorbed by the receiving array, thereby obtaining the original ultrasonic field data. The calculation formula is as follows:

[0059] ;

[0060] in, This represents the obtained raw ultrasonic field data; Indicates the relative angle of the scanned data; This represents the signal amplitude after projection transformation; y represents the position in the original spatial plane; A function representing the geometric shape of the object being measured.

[0061] Specifically, methods for arranging the raw ultrasonic field data by angular direction and then filtering include: mean filtering, first-order filtering, or Gaussian filtering.

[0062] Specifically, in this embodiment, the ultrasonic array is placed horizontally or vertically, and the workpiece to be tested is placed on top of the linear module and attracted by a magnet. The motor performs circular motion to drive the linear module to perform linear or rotational motion. The wave function signals used include square wave signals or sine wave signals, and the filtered ultrasonic field data are arranged into a two-dimensional matrix. Each dimension of the two-dimensional matrix represents the actual array element arrangement order, and the complex values ​​in the two-dimensional matrix... This represents the amplitude and phase information of the ultrasonic signal emitted by array element i and received at array element j.

[0063] S2, the original ultrasonic field data is arranged according to the angular direction and then filtered to obtain the filtered signal matrix. Based on the signal matrix, the original ultrasonic field data is transformed in the frequency domain to obtain the frequency domain data of the original ultrasonic field data. The frequency domain data is filtered twice using a ramp function to obtain the signal matrix after secondary filtering. The signal matrix after secondary filtering is spliced ​​and inversely transformed to obtain a low-resolution image.

[0064] Specifically, the formula for calculating low-resolution images is as follows:

[0065] ;

[0066] ;

[0067] ;

[0068] in, This represents the obtained raw ultrasonic field data; Represents the frequency domain transform function; Indicates the window function; Indicates angular frequency; and These represent the components of the angular frequency on the horizontal axis and the components on the vertical axis, respectively. represents an imaginary number; This represents the frequency domain function after ramp filtering; Represents the real number field.

[0069] S3: Obtain and perform preset size adjustment and equal-interval segmentation on each slice data image in the open-source high-resolution CT dataset NIST to obtain feature blocks that constitute the minimum elements of the external dictionary. Cluster the feature blocks that constitute the minimum elements of the external dictionary to obtain several feature groups, which constitute the external dictionary. Perform preset size adjustment and equal-interval segmentation on the obtained low-resolution image to obtain feature blocks that constitute the minimum elements of the internal dictionary. Directly use the feature blocks that constitute the minimum elements of the internal dictionary as dictionary elements to obtain the internal dictionary.

[0070] Specifically, the construction methods for external and internal dictionaries are as follows:

[0071] For each slice of data in the open-source high-resolution CT dataset The image size was adjusted to 256*256 pixels. Each slice was divided into 1024 8*8 pixel feature blocks. These feature blocks were then clustered using the KNN algorithm to obtain j feature groups, with an external dictionary. Represented as ,in This represents the feature block that constitutes the smallest element of the external dictionary. This represents several feature groups after clustering.

[0072] The internal dictionary V generated from the low-resolution image contains only one data slice. Data slicing After resizing the image to 256*256, it was segmented at equal intervals, resulting in 1024 8*8 feature blocks. The internal dictionary... Represented as , This represents the feature block that constitutes the smallest element of the internal dictionary.

[0073] S4. Based on the constructed internal and external dictionaries, an optimization function is used to couple each feature block in the internal and external dictionaries with the low-resolution image at the time and frequency domain levels to obtain coupled dictionary features. The mapping relationship of the coupled dictionary features is optimized. Based on the optimized mapping relationship and the coupling loss function, the FISTA algorithm is used to iterate the coupled dictionary features to obtain the optimal mapping relationship between dictionary elements and the low-resolution image. When the coupling loss function reaches its minimum through iteration, the obtained optimal mapping relationship is used to reconstruct the low-resolution image based on compressed sensing to obtain the reconstructed tomographic image.

[0074] Specifically, optimization functions are used to couple each feature block in the internal and external dictionaries with the low-resolution image at both the time and frequency domain levels to obtain coupled dictionary features. The calculation formula is as follows:

[0075]

[0076] ;

[0077] in, Dictionary features indicating coupling; , and Indicates hyperparameters; Indicates an external dictionary; Indicates an internal dictionary; and This indicates that a frequency domain transformation is performed on the feature groups in the dictionary; This represents the frequency domain transform function.

[0078] Specifically, in S4, the calculation formula of the FISTA algorithm is as follows:

[0079] ;

[0080] ;

[0081] ;

[0082] in, This represents the intermediate value of the mapping relationship after the k-th round of optimization; Indicates the step size adjustment factor; This represents the final value of the mapping relationship after the k-th round of optimization; Indicates hyperparameters; ( ) indicates that the coupled dictionary features are transposed; Multiplying the dictionary matrix by the mapping relationship of the (k-1)th round, we get the intermediate state image matrix of the (k-1)th round. This indicates a low-resolution image.

[0083] S5 stitches the reconstructed tomographic images in height order to obtain a high-resolution three-dimensional image of the workpiece under test.

[0084] Specifically, the reconstructed tomographic images are stitched together in height order, as follows:

[0085] The current position of the workpiece is marked as the original position; the unit distance of each workpiece translation is recorded, and the imaging data before and after a single translation is compared by a time-series algorithm to check whether the reconstructed image after translation includes the internal information of the workpiece; if it is still within the imaging range of the workpiece, the reconstructed image is stitched together at a unit distance in the third axis direction; otherwise, the stitching is terminated to obtain the final reconstructed tomographic image.

[0086] Specifically, in this embodiment, the final obtained three-dimensional image contains the outline and internal structure information of the object being tested, and the internal structure of the object can be judged by the naked eye or intelligent algorithms to determine whether it meets the production standards.

[0087] Specifically, in this embodiment, each feature block in the recovered high-resolution image The specific method can be found in the following formula:

[0088] ;

[0089] Where y represents the degraded image, D represents the optimized dictionary, and F(.) represents the frequency domain transform. , , This represents the hyperparameter, and A represents the high-resolution image.

[0090] Specifically, after dictionary optimization, the FISTA method is used to iterate over the degraded image. The iteration process is represented as a Sub-A subproblem, and its specific optimization formula is as follows:

[0091] ;

[0092] ;

[0093] Soft is a soft threshold function, which is expressed as follows during the calculation:

[0094] .

[0095] Specifically, in this embodiment, a detection system is used, comprising a detection water tank, a workpiece clamping and displacement module, and an ultrasonic emission and acquisition module (corresponding to an ultrasonic array). The detection water tank provides the detection environment, the workpiece clamping and displacement module displaces the workpiece to obtain data at different angles and heights, and the ultrasonic emission and acquisition module transmits and receives ultrasonic data to obtain data that meets the conditions for image inversion. The detection water tank ensures that the workpiece clamping and displacement module and the ultrasonic acquisition module are completely submerged in liquid to meet the requirements of ultrasonic detection. The workpiece clamping and displacement module consists of a stepper motor and a linear module, with the stepper motor shaft, the center of the linear module, and the geometric center of the clamped workpiece coaxial. The workpiece is placed on top of the linear module and attracted by a magnet. Displacement is achieved by the stepper motor shaft's circular motion driving the linear module's output components to move linearly and rotate. The ultrasonic emission and acquisition module consists of an ultrasonic transducer, including an ultrasonic transmitter and an ultrasonic receiver. High-frequency vibrations of a crystal oscillator in the ultrasonic transmitter generate ultrasonic signals. Part of the signal passes through the workpiece, and part is reflected by the workpiece. The wave field signals are received by the ultrasonic receivers on both sides.

[0096] The specific detection steps of the detection system described in this embodiment are as follows: Liquid is poured into the detection tank to submerge the ultrasonic probe, satisfying the conditions for ultrasonic propagation. The workpiece to be tested is placed on top of the workpiece clamping and displacement module and attracted by a magnet, ensuring that the geometric center of the workpiece, the center of rotation of the linear module, and the motor shaft are coaxial. The stepper motor rotates at a constant speed according to the input pulse duty cycle, driving the linear module to perform helical motion along the Z-axis. Simultaneously with the workpiece displacement, the ultrasonic transmission and acquisition module emits ultrasonic signals, and the other receivers receive the echo data. The received data is processed and then subjected to full waveform inversion to achieve ultrasonic tomographic imaging of the tested object. The generated two-dimensional tomographic images are stitched together to obtain a final three-dimensional internal scan of the tested workpiece.

[0097] The ultrasonic probe is placed on the platform of the testing tank and fixed with bolts. Its relative distance can be adjusted via knobs on both sides. A single ultrasonic probe comprises 64 array elements. One element is controlled to emit ultrasonic waves at a time, while the remaining elements act as receivers to receive the echo data. The data obtained from one transmission and reception cycle can be used to invert and obtain the internal sound velocity distribution of a vertical cross-section of the workpiece within the scanning plane. After receiving the echo signal from a vertical cross-section of the workpiece, a stepper motor is controlled to advance by a fixed angle, causing the workpiece to rotate clockwise by a fixed angle. After rotating to the required angle, the ultrasonic probe repeats the transmission and reception steps, obtaining tomographic ultrasonic data for each scanning cross-section by continuously rotating and ascending the workpiece.

[0098] The data post-processing specifically involves: arranging the echo data received at the same time into a complex observation matrix according to the transmission and reception relationship of the array elements. The element in the i-th row and j-th column of the matrix represents the wave field signal emitted by the j-th transmitting source and received by the i-th receiver. Then, the wave field matrix is ​​filtered to remove low-frequency noise. The filtered signal is input into a full waveform inversion algorithm, using the finite difference method to generate the model and invert the parameters. Finally, a matrix inversion generates the internal sound velocity distribution of the workpiece in a vertical section of the scanning plane. The complex observation matrices at different times correspond to different scanning sections of the workpiece under test. After image alignment, three-dimensional reconstruction is performed, followed by color mapping, ultimately obtaining a three-dimensional sound velocity image of the workpiece under test. The ultrasonic probe (i.e., the ultrasonic array) can be placed horizontally or vertically. The more ultrasonic array elements (the ultrasonic array includes ultrasonic array elements) on the same plane, the higher the accuracy of the final inverted image. The ultrasonic array element transmission and reception methods can be combined in different ways. Two schemes are: one array element transmits a signal at a time, and all other array elements on both sides receive the signal; the other array element transmits a signal at a time, and all array elements on the opposite side receive the signal. The ultrasonic array element method selects window functions such as Ricker wavelet or Hanning window for excitation signals, analyzes the signals by converting them to the frequency domain, and optimizes the full waveform inversion using a prior initial acoustic model. Specifically, it designs an initial model that conforms to the real sound velocity distribution based on the sound velocity propagation characteristics of the workpiece as the first-round forward model of the full waveform inversion algorithm to accelerate the generation of inversion images.

[0099] In summary, this embodiment achieves 3D image reconstruction by controlling the workpiece displacement and installing only array elements on both sides of the linear array. In the 3D image reconstruction, this invention references a 2D ultrasonic image reconstruction method. By controlling the array elements to emit and capture ultrasonic waves at each horizontal cross-section during the workpiece's uniform rotation and ascent, an ultrasonic tomographic image of each cross-section is obtained. After the 2D image inversion is complete, the inverted images are stitched together in the third dimension to obtain the desired 3D ultrasonic image of the workpiece. The principle of this invention can be understood as follows: The motion module controls the vertical and rotational movement of the workpiece. Simultaneously with each unit height as the workpiece spirals upward, each emission unit on the linear array sequentially emits ultrasonic signals, and the received signals are imaged to obtain the scanning result of the current plane. Finally, the images of each plane are stitched together to obtain the final 3D imaging image.

[0100] Specifically, in this embodiment, each image slice ensures that the scanning plane of the ultrasound array is submerged in water during the test environment. The imaging quality of each slice is evaluated using the structural similarity index SSIM and peak signal-to-noise ratio PSNR, and the average value of all slices is taken as the evaluation result of the three-dimensional image. Figure 2As shown, in this embodiment, low-resolution ultrasound data is acquired through a hardware system, and a dictionary of high- and low-resolution image patches and their mapping relationships are constructed by combining them with an external standard dataset. Subsequently, the mapping relationship is used to guide the deep neural network to perform iterative training and feature mapping, reconstructing the input low-resolution image into a high-resolution image. Finally, the reconstructed continuous slice images are stacked to generate a three-dimensional visualization model of the target.

[0101] like Figure 3 As shown, this embodiment also discloses a dictionary-based 3D image reconstruction system, including:

[0102] The ultrasonic field data acquisition module 31 is used to acquire the original ultrasonic field data generated by the workpiece under test during the rotational scanning process.

[0103] The low-resolution image acquisition module 32 is used to arrange the original ultrasonic field data according to the angular direction and then filter it to obtain the filtered signal matrix. Based on the signal matrix, the original ultrasonic field data is transformed in the frequency domain to obtain the frequency domain data of the original ultrasonic field data. The frequency domain data is filtered twice using a ramp function to obtain the signal matrix after secondary filtering. The signal matrix after secondary filtering is spliced ​​and inversely transformed to obtain a low-resolution image.

[0104] The internal and external dictionary generation module 33 is used to acquire and perform preset size adjustment and equal-interval segmentation on each slice data image in the open-source high-resolution CT dataset NIST to obtain feature blocks that constitute the minimum element of the external dictionary. The feature blocks that constitute the minimum element of the external dictionary are clustered to obtain several feature groups, which constitute the external dictionary. The module also performs preset size adjustment and equal-interval segmentation on the obtained low-resolution image to obtain feature blocks that constitute the minimum element of the internal dictionary. The module directly uses the feature blocks that constitute the minimum element of the internal dictionary as dictionary elements to obtain the internal dictionary.

[0105] The tomographic image reconstruction module 34 is used to couple each feature block in the internal and external dictionaries with the low-resolution image at the time and frequency domain levels, respectively, based on the constructed internal and external dictionaries, to obtain coupled dictionary features. The mapping relationship of the coupled dictionary features is optimized, and the FISTA algorithm is used to iterate the coupled dictionary features based on the optimized mapping relationship and the coupling loss function to obtain the best mapping relationship between dictionary elements and the low-resolution image. When the coupling loss function reaches its minimum through iteration, the obtained best mapping relationship is used to reconstruct the low-resolution image based on compressed sensing to obtain the reconstructed tomographic image.

[0106] The high-resolution image acquisition module 35 is used to stitch the reconstructed tomographic images in height order to obtain a high-resolution three-dimensional image of the workpiece under test.

Claims

1. A three-dimensional image reconstruction method based on dictionary learning, characterized in that, include: S1, acquire the original ultrasonic field data generated by the workpiece under test during the rotational scanning process; S2, the original ultrasonic field data is arranged according to the angular direction and the first filtering is performed to obtain the first signal matrix. Based on the first signal matrix, the original ultrasonic field data is transformed in the frequency domain to obtain the frequency domain data of the original ultrasonic field data. The frequency domain data is filtered twice using the ramp function to obtain the second signal matrix. The second signal matrix is ​​stitched together and inversely transformed to obtain a low-resolution image. S3: Obtain each slice image from the open-source high-resolution CT dataset NIST and perform preset size adjustment and equal-interval segmentation to obtain feature blocks that constitute the minimum elements of the outer dictionary. Cluster the feature blocks that constitute the minimum elements of the outer dictionary to obtain several feature groups, which constitute the outer dictionary. Perform preset size adjustment and equal-interval segmentation on the obtained low-resolution image to obtain feature blocks that constitute the minimum elements of the inner dictionary. Use the feature blocks that constitute the minimum elements of the inner dictionary as dictionary elements to obtain the inner dictionary. S4. Based on the internal and external dictionaries, an optimization function is used to couple each feature block in the internal and external dictionaries with the low-resolution image at both the time and frequency domain levels to obtain coupled dictionary features. The mapping relationship of the coupled dictionary features is optimized using an optimization function that correlates the time and frequency domains. Based on the optimized mapping relationship and the coupling loss function, the FISTA algorithm is used to iterate the coupled dictionary features to obtain the optimal mapping relationship between dictionary elements and the low-resolution image. When the coupling loss function reaches its minimum through iteration, the obtained optimal mapping relationship is used to reconstruct the low-resolution image based on compressed sensing to obtain the reconstructed tomographic image. The optimization function is used to couple each feature block in the internal and external dictionaries with the low-resolution image at both the time and frequency domain levels to obtain the coupled dictionary features. The calculation formula is as follows: ； in, Dictionary features indicating coupling; , and Indicates hyperparameters; Indicates an external dictionary; Indicates an internal dictionary; and This indicates that a frequency domain transformation is performed on the feature groups in the dictionary; Represents the frequency domain transform function; S5 stitches the reconstructed tomographic images in height order to obtain a high-resolution three-dimensional image of the workpiece under test.

2. The dictionary-based 3D image reconstruction method according to claim 1, characterized in that, In S1, the process of acquiring the original ultrasonic field data specifically includes: S11, Pour liquid into the test tank to cover the ultrasonic array, align the geometric center of the workpiece under test with the rotation center of the linear module, and make the geometric center of the workpiece under test, the rotation center of the linear module, and the center of the motor shaft on the same axis. S12, the transmitting array receives the transmission sequence signal and wave function signal transmitted by the control center, the receiving array receives the receiving sequence signal transmitted by the control center, and the motor driver receives the motor pulse ratio signal transmitted by the control center. Based on the transmission sequence signal, wave function signal, receiving sequence signal and motor pulse ratio signal, the drive motor rotates by one step angle and a set of ultrasonic signals penetrates the workpiece under test and is absorbed by the receiving array, thereby obtaining the original ultrasonic field data.

3. The dictionary-based 3D image reconstruction method according to claim 2, characterized in that, Each time the drive motor rotates by one step angle, a set of ultrasonic signals penetrates the workpiece under test and is absorbed by the receiving array, thereby obtaining the original ultrasonic field data. The calculation formula is as follows: ； in, This represents the obtained raw ultrasonic field data; Indicates the relative angle of the scanned data; This represents the signal amplitude after projection transformation; y represents the position in the original spatial plane; A function representing the geometric shape of the object being measured.

4. The dictionary-based 3D image reconstruction method according to claim 1, characterized in that, In S2, the methods for arranging the raw ultrasonic field data by angular direction and then filtering include: mean filtering, first-order filtering, or Gaussian filtering.

5. The dictionary-based 3D image reconstruction method according to claim 3, characterized in that, In S2, the formula for calculating low-resolution images is as follows: ； ； ； in, This represents the obtained raw ultrasonic field data; Represents the frequency domain transform function; Indicates the window function; Indicates angular frequency; and These represent the components of the angular frequency on the horizontal axis and the components on the vertical axis, respectively. represents an imaginary number; This represents the frequency domain function after ramp filtering; Represents the real number field.

6. The dictionary-based 3D image reconstruction method according to claim 1, characterized in that, In S3, the construction methods for external and internal dictionaries are as follows: For each slice of data in the open-source high-resolution CT dataset The image size was adjusted to 256*256 pixels. Each slice was divided into 1024 8*8 pixel feature blocks. These feature blocks were then clustered using the KNN algorithm to obtain j feature groups, with an external dictionary. Represented as ,in This represents the feature block that constitutes the smallest element of the external dictionary. Represents several feature groups after clustering; The internal dictionary V generated from the low-resolution image contains only one data slice. Data slicing After resizing the image to 256*256, it was segmented at equal intervals, resulting in 1024 8*8 feature blocks. The internal dictionary... Represented as , This represents the feature block that constitutes the smallest element of the internal dictionary.

7. The dictionary-based 3D image reconstruction method according to claim 1, characterized in that, In S4, the calculation formula for the FISTA algorithm is as follows: ； ； ； in, This represents the intermediate value of the mapping relationship after the k-th round of optimization; Indicates the step size adjustment factor; This represents the final value of the mapping relationship after the k-th round of optimization; Indicates hyperparameters; ( ) indicates that the coupled dictionary features are transposed; Multiplying the dictionary matrix by the mapping relationship of the (k-1)th round, we get the intermediate state image matrix of the (k-1)th round. This indicates a low-resolution image.

8. The dictionary-based 3D image reconstruction method according to claim 1, characterized in that, In S5, the reconstructed tomographic images are stitched together in height order, specifically as follows: The current position of the workpiece is marked as the original position; the unit distance of each workpiece translation is recorded, and the imaging data before and after a single translation is compared by a time-series algorithm to check whether the reconstructed image after translation includes the internal information of the workpiece; if it is still within the imaging range of the workpiece, the reconstructed image is stitched together at a unit distance in the third axis direction; otherwise, the stitching is terminated to obtain the final reconstructed tomographic image.

9. A three-dimensional image reconstruction system based on dictionary learning, characterized in that, include: The ultrasonic field data acquisition module is used to acquire the raw ultrasonic field data generated by the workpiece under test during the rotational scanning process; The low-resolution image acquisition module is used to arrange the original ultrasonic field data according to the angular direction and then filter it to obtain the filtered signal matrix. Based on the signal matrix, the original ultrasonic field data is transformed in the frequency domain to obtain the frequency domain data of the original ultrasonic field data. The frequency domain data is filtered twice using a ramp function to obtain the signal matrix after secondary filtering. The signal matrix after secondary filtering is stitched together and inversely transformed to obtain the low-resolution image. The internal and external dictionary generation modules are used to acquire and perform preset size adjustment and equal-interval segmentation on each slice data image in the open-source high-resolution CT dataset NIST to obtain feature blocks that constitute the minimum elements of the external dictionary. The feature blocks that constitute the minimum elements of the external dictionary are clustered to obtain several feature groups, which constitute the external dictionary. The modules also perform preset size adjustment and equal-interval segmentation on the obtained low-resolution images to obtain feature blocks that constitute the minimum elements of the internal dictionary. The feature blocks that constitute the minimum elements of the internal dictionary are directly used as dictionary elements to obtain the internal dictionary. The tomographic image reconstruction module uses an optimization function to couple each feature block in the internal and external dictionaries with the low-resolution image at both the time and frequency domain levels, based on the constructed internal and external dictionaries, to obtain coupled dictionary features. The mapping relationship of the coupled dictionary features is then optimized. Based on the optimized mapping relationship and the coupling loss function, the FISTA algorithm is used to iterate the coupled dictionary features to obtain the optimal mapping relationship between dictionary elements and the low-resolution image. When the coupling loss function reaches its minimum through iteration, the obtained optimal mapping relationship is used to reconstruct the low-resolution image based on compressed sensing, resulting in the reconstructed tomographic image. The optimization function is used to couple each feature block in the internal and external dictionaries with the low-resolution image at both the time and frequency domain levels to obtain the coupled dictionary features. The calculation formula is as follows: ； in, Dictionary features indicating coupling; , and Indicates hyperparameters; Indicates an external dictionary; Indicates an internal dictionary; and This indicates that a frequency domain transformation is performed on the feature groups in the dictionary; Represents the frequency domain transform function; The high-resolution image acquisition module is used to stitch together the reconstructed tomographic images in height order to obtain a high-resolution three-dimensional image of the workpiece under test.

Images