A Regional Infrared Digital Holography Method Based on Convolutional Neural Networks

Abstract

The invention discloses a regional infrared digital holographic method based on a convolutional neural network, including infrared digital holographic system construction and data set acquisition, simulated sample data set construction, dynamic reconstruction based on a convolutional neural network model, and regional block reconstruction of the original infrared hologram. Its advantage is that the original infrared hologram containing multiple samples at different depths is converted into an infrared digital holographic reconstructed amplitude and phase image with different focusing distances in different image regions.

Description

Technical Field

[0001] The invention relates to the technical field of infrared digital holographic imaging, and in particular to a regional infrared digital holographic method based on a convolutional neural network. Background Art

[0002] Compared with digital holography in the visible light band, digital holography in the infrared band has different penetration into samples to obtain different holographic images, which makes it have broad application prospects in metal flaw detection and biomedical testing. Convolutional neural networks appeared in the 1980s. Now, with the continuous development of deep learning theory and the continuous improvement of CPU / GPU equipment, the situation that convolutional neural networks were limited by computing power in the last century has been solved and developed rapidly, showing its powerful ability in applications in fields such as computer vision and natural language processing. In traditional imaging methods, for the same reconstructed amplitude and phase image, only one focusing distance can be selected for reconstruction. When there are multiple samples at different positions in the image, the final amplitude and phase images are of poor quality. Combining the unique advantages of convolutional neural networks, a regional infrared digital holography method based on convolutional neural networks can be realized.

[0003] The present invention aims to propose a regional infrared digital holography method based on a convolutional neural network. The method belongs to a new regional infrared wave digital holography imaging technology. The existing regional infrared wave digital holography imaging method has the following shortcomings: for the same reconstructed amplitude and phase image, only one focusing distance can be selected for reconstruction. When there are multiple samples at different positions in the image, the final amplitude and phase image quality is poor. The method proposed by the present invention can transform the original infrared hologram containing multiple samples at different depths into an infrared digital holographic reconstructed amplitude and phase image with different focusing distances in different areas of the image through an algorithm given by a model. Summary of the invention

[0004] The purpose of an embodiment of the present invention is to provide a regional infrared digital holography method based on a convolutional neural network, so as to solve the problem that for the same reconstructed amplitude and phase image, only one focusing distance can be selected for reconstruction, and when there are multiple samples at different positions in the image, the final amplitude and phase image quality is poor.

[0005] To achieve the above objectives, the present invention provides the following technical solutions to solve the above problems:

[0006] A regional infrared digital holographic method based on a convolutional neural network comprises the following steps:

[0007] Step 1: Infrared digital holographic system construction and data set acquisition process;

[0008] The infrared digital holographic system comprises: a laser light source 1, a collimator 2, a light beam splitter 3, a reflector 1 4, a reflector 2 5, a two-dimensional electric translation stage 6, an axial translation stage 7, a sample 8, a connecting rod 9, and an infrared detector 10;

[0009] The laser light source 1 is used to emit infrared laser and is connected to the collimator 2 through an optical fiber. The infrared laser emitted from the collimator 2 is divided into two coherent lights after passing through the beam splitter 3. One infrared laser passes through the sample 8 and then passes through the reflector 1 4 located behind the axial translation stage 7 to reach the infrared detector 10, and the other infrared laser passes through the reflector 2 5 and directly reaches the infrared detector 10. The two-dimensional electric translation stage 6 and the axial translation stage 7 are connected to each other using a connecting rod 9 so that the sample can be translated or rotated; the collimator 2, the beam splitter 3, the axial translation stage 7 and the reflector 1 4 are arranged vertically in space in sequence; the infrared detector 10 is parallel to the reflector 1 4; the infrared detector 10 is used to receive infrared laser signals and is located directly below the reflector 2 5; the reflector 2 5 is parallel to the beam splitter 3;

[0010] The original infrared hologram is obtained through the infrared digital holography system. The glass slide covered with the sample is placed on the axial translation stage. The sample is continuously moved in the axial direction, and the original infrared holograms of multiple axial depths are collected by the infrared detector. The collected original infrared holograms are then reconstructed through the algorithm to obtain the real sample data set. The focal distance obtained from the axial translation stage is recorded to calibrate the real sample data set.

[0011] Step 2: Simulate the sample data set construction process. Use the free space diffraction propagation algorithm in the computer to simulate the hologram focal plane reconstruction process. First, generate the original simulated hologram of the random library sample, simulate and calculate the diffraction image obtained after free space propagation at different distances, and manually calibrate the focusing effect of each diffraction image. Finally, the original simulated hologram and the diffraction image form a simulated sample data set and divide it into a training set and a test set according to the ratio of M:N. Here, the automatic focusing method is used to simulate the generation of the reconstruction distance, reconstruct the image on the plane within the entire propagation distance range, and use the numerical results of different propagation distances to generate multiple simulated focused holograms.

[0012] Step 3: Based on the dynamic reconstruction process of the convolutional neural network model, identify the number of samples that need to be focused during the reconstruction process, pre-process the real sample data set through a computer, identify the number of samples corresponding to each original infrared hologram in the real sample data set, and calibrate the range of each sample to generate a parameter file corresponding to the real sample data set. Establish a convolutional neural network model, use the training set and test set obtained in step 2, set the neural network parameters to extract the features of holograms with different focusing degrees, set the back propagation process, and use the focusing distance in the real sample data set calibrated in step 1 to establish the loss function.

[0013] The original infrared hologram is propagated to K planes at different distances using the angular spectrum algorithm within the range of K prediction results. Each original infrared hologram obtained through the reconstruction process is used to construct U original infrared focused holograms in the form of slice segmentation focusing distance for screening, and compared with the initial distance when the original infrared hologram was recorded in step one for verification. A secondary loss function is constructed to eliminate errors and finally output the predicted optimal focusing distance. Training is performed until the training results in the test set and the calibrated real sample data set have a focusing distance accuracy of more than P. The real sample data set is input into this convolutional neural network model after reading the parameter file to perform free space propagation reconstruction on the real sample data set.

[0014] Step 4: The original infrared hologram regional block reconstruction process, the original infrared hologram is gridded and divided into m*n infrared regional holograms, the convolutional neural network model in step 3 is used to determine the focusing distance of each infrared regional hologram and output to obtain a parameter data set, the parameter data set contains the optimal focusing distance of each infrared regional hologram in different ranges in the real sample data set. Finally, the focusing method in step 2 is applied to the real sample data set and the parameter data set to obtain the infrared digital holographic block reconstruction image, and finally the m*n infrared digital holographic block reconstruction images are spliced ​​according to the gridded block model and the sample edge outer background between each infrared digital holographic block reconstruction image and its adjacent infrared digital holographic block reconstruction image is reconstructed using the same focusing distance to prevent gaps between the spliced ​​images to achieve the purpose of smooth splicing, and finally the infrared digital holographic reconstruction amplitude and phase images are obtained.

[0015] Preferably, the loss function used in step three is an adjustable function.

[0016] The embodiments of the present invention have the following advantages:

[0017] Technically, the original infrared hologram containing multiple samples at different depths is transformed into an infrared digital holographic reconstructed amplitude and phase image with different focusing distances in different areas of the image. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a flow chart of a regional infrared digital holographic method based on a convolutional neural network of the present invention;

[0019] Figure 2 This is a structural diagram of the imaging system of a regional infrared digital holography method based on a convolutional neural network in the present invention. DETAILED DESCRIPTION

[0020] The following examples are used to illustrate the present invention but are not intended to limit the scope of the present invention. Example

[0021] Please refer to the content in the figure. The present invention provides a regional infrared digital holographic method based on a convolutional neural network, which is characterized by: Figure 1 The following steps are shown:

[0022] Step 1: Infrared digital holographic system construction and data set acquisition process;

[0023] like Figure 2 As shown, the infrared digital holographic system includes: a laser light source 1, a collimator 2, a light beam splitter 3, a reflector 1 4, a reflector 2 5, a two-dimensional electric translation stage 6, an axial translation stage 7, a sample 8, a connecting rod 9, and an infrared detector 10;

[0024] The laser light source 1 is used to emit infrared laser and is connected to the collimator 2 through an optical fiber. The infrared laser emitted from the collimator 2 is divided into two coherent lights after passing through the beam splitter 3. One infrared laser passes through the sample 8 and then passes through the reflector 1 4 located behind the axial translation stage 7 to reach the infrared detector 10, and the other infrared laser passes through the reflector 2 5 and directly reaches the infrared detector 10. The two-dimensional electric translation stage 6 and the axial translation stage 7 are connected to each other using a connecting rod 9 so that the sample can be translated or rotated; the collimator 2, the beam splitter 3, the axial translation stage 7 and the reflector 1 4 are arranged vertically in space in sequence; the infrared detector 10 is parallel to the reflector 1 4; the infrared detector 10 is used to receive infrared laser signals and is located directly below the reflector 2 5; the reflector 2 5 is parallel to the beam splitter 3;

[0025] The original infrared hologram is obtained through the infrared digital holography system. The glass slide covered with the sample is placed on the axial translation stage. The sample is continuously moved in the axial direction, and the original infrared holograms of multiple axial depths are collected by the infrared detector. The collected original infrared holograms are then reconstructed through the algorithm to obtain the real sample data set. The focal distance obtained from the axial translation stage is recorded to calibrate the real sample data set.

[0026] Step 2: Simulate the sample data set construction process; Use the free space diffraction propagation algorithm in the computer to simulate the hologram focal plane reconstruction process. First, generate the original simulated hologram of the random library sample, simulate and calculate the diffraction image obtained after free space propagation at different distances, and manually calibrate the focusing effect of each diffraction image. Finally, the original simulated hologram and the diffraction image form a simulated sample data set and divide it into a training set and a test set at a ratio of 1000:1. Here, the automatic focusing method is used to simulate the generation of the reconstruction distance, reconstruct the image on the plane within the entire propagation distance range, and use the numerical results of different propagation distances to generate multiple simulated focused holograms.

[0027] Step 3: Dynamic reconstruction process based on convolutional neural network model; identify the number of samples that need to be focused during the reconstruction process, pre-process the real sample data set through a computer, identify the number of samples corresponding to each original infrared hologram in the real sample data set, and calibrate the range of each sample to generate a parameter file corresponding to the real sample data set. Establish a convolutional neural network model, use the training set and test set obtained in step 2, set the neural network parameters to extract the features of holograms with different focusing degrees, set the back propagation process, and use the focusing distance in the real sample data set calibrated in step 1 to establish the loss function.

[0028] The original infrared hologram is propagated to 10 planes at different distances using the angular spectrum algorithm within the range of 10 prediction results. Each original infrared hologram obtained through the reconstruction process is used to construct 100 original infrared focused holograms in the form of slice segmentation focusing distance for screening, and compared with the initial distance when the original infrared hologram was recorded in step one. A secondary loss function is constructed to eliminate errors and finally output the predicted optimal focusing distance. Training is performed until the training results in the test set and the calibrated real sample data set have a focusing distance accuracy of more than 99.5%. The real sample data set is input into this convolutional neural network model after reading the parameter file to perform free space propagation reconstruction on the real sample data set.

[0029] Step 4, the regional block reconstruction process of the original infrared hologram; the original infrared hologram containing 4 samples is gridded and divided into 2*2 infrared regional holograms, and the convolutional neural network model in step 3 is used to determine the focusing distance of each infrared regional hologram and output to obtain a parameter data set, which contains the optimal focusing distance of each infrared regional hologram in different ranges in the real sample data set. Finally, the focusing method in step 2 is applied to the real sample data set and the parameter data set to obtain the infrared digital holographic block reconstruction image, and finally the 2*2 infrared digital holographic block reconstruction images are spliced ​​according to the gridded block model, and the background outside the sample edge between each infrared digital holographic block reconstruction image and its adjacent infrared digital holographic block reconstruction image is reconstructed using the same focusing distance to prevent gaps between the spliced ​​images to achieve the purpose of smooth splicing, and finally the infrared digital holographic reconstruction amplitude and phase images are obtained.

[0030] The experimental results of the typical embodiments of the present invention prove that the invention can effectively solve the problem that only one focusing distance can be selected for reconstruction, and when there are multiple samples at different positions in the image, the resulting amplitude and phase images are of poor quality. Technically, the original infrared hologram containing multiple samples at different depths is converted into an infrared digital holographic reconstructed amplitude and phase image with different focusing distances in different areas of the image.

[0031] Although the present invention has been described in detail above by general description and specific embodiments, it is obvious to those skilled in the art that some modifications or improvements can be made to the present invention. Therefore, these modifications or improvements made without departing from the spirit of the present invention all belong to the scope of protection claimed by the present invention.

Claims

1. A regional infrared digital holographic method based on convolutional neural network, characterized in that: The following steps are involved: Step 1: Infrared digital holographic system construction and data set acquisition process; The infrared digital holographic system comprises: a laser light source (1), a collimator (2), a light beam splitter (3), a reflector 1 (4), a reflector 2 (5), a two-dimensional electric translation stage (6), an axial translation stage (7), a sample (8), a connecting rod (9), and an infrared detector (10); The laser light source (1) is used to emit infrared laser light and is connected to the collimator (2) through an optical fiber. The infrared laser light emitted from the collimator (2) is split into two coherent lights after passing through a beam splitter (3). One of the infrared laser lights passes through a sample (8) and then passes through a reflector 1 (4) located behind an axial translation stage (7) to reach an infrared detector (10). The other infrared laser light passes through a reflector 2 (5) and then directly reaches the infrared detector (10). The two-dimensional electric translation stage (6) and the axial translation stage (7) are connected to each other using a connecting rod (9) so that the sample can be translated or rotated. The collimator (2), the beam splitter (3), the axial translation stage (7) and the reflector 1 (4) are arranged vertically in space in sequence. The infrared detector (10) is parallel to the reflector 1 (4). The infrared detector (10) is used to receive infrared laser signals and is located directly below the reflector 2 (5). The reflector 2 (5) is parallel to the beam splitter (3). Acquire an original infrared hologram through an infrared digital holographic system, place a glass slide covered with a sample (8) above an axial translation stage (7), continuously move the sample (8) in the axial direction, and acquire original infrared holograms at multiple axial depths using an infrared detector (10), and then reconstruct the acquired original infrared holograms through an algorithm to obtain a real sample data set; record the focal distance obtained from the axial translation stage (7), and calibrate the real sample data set; Step 2: simulate the sample data set construction process, use the free space diffraction propagation algorithm in the computer to simulate the hologram focal plane reconstruction process, first generate the original simulated hologram of the random library sample, simulate and calculate the diffraction image obtained after free space propagation at different distances, and manually calibrate the focusing effect of each diffraction image, and finally form a simulated sample data set composed of the original simulated hologram and the diffraction image and divide it into a training set and a test set according to the ratio of M:N; here, the automatic focusing method is used to simulate the generation of the reconstruction distance, reconstruct the image on the plane within the entire propagation distance range, and use the numerical results of different propagation distances to generate multiple simulated focused holograms; Step 3: Based on the dynamic reconstruction process of the convolutional neural network model, identify the number of samples that need to be focused during the reconstruction process, pre-process the real sample data set through a computer, identify the number of samples corresponding to each original infrared hologram in the real sample data set, and calibrate the range of each sample to generate a parameter file corresponding to the real sample data set; establish a convolutional neural network model, set the neural network parameters to extract the features of holograms with different focusing degrees by using the training set and test set obtained in step 2, and set a back propagation process, and use the focusing distance in the real sample data set calibrated in step 1 to establish a loss function; Using the angular spectrum algorithm within the range of K prediction results, the original infrared hologram is propagated to K planes at different distances. Each original infrared hologram obtained through the reconstruction process is used to construct U original infrared focused holograms in the form of slice segmentation focus distance for screening, and compared with the initial distance when the original infrared hologram was recorded in step 1 for verification, and a secondary loss function is constructed to eliminate errors and finally output the predicted optimal focus distance. Training is performed until the training results in the test set and the focus distance accuracy of the calibrated real sample data set reach P or more, and the real sample data set is input into this convolutional neural network model after reading the parameter file to reconstruct the real sample data set through free space propagation; Step 4, the regional block reconstruction process of the original infrared hologram, the original infrared hologram is gridded and blocked to obtain m*n infrared regional holograms, the convolutional neural network model in step 3 is used to determine the focusing distance of each infrared regional hologram and output to obtain a parameter data set, the parameter data set contains the optimal focusing distance of each infrared regional hologram in different ranges in the real sample data set; finally, the focusing method in step 2 is applied to the real sample data set and the parameter data set to obtain the infrared digital holographic block reconstructed image, and finally the m*n infrared digital holographic block reconstructed images are spliced ​​according to the gridded block model and the background outside the sample edge between each infrared digital holographic block reconstructed image and its adjacent infrared digital holographic block reconstructed image is reconstructed using the same focusing distance to prevent gaps between the spliced ​​images to achieve the purpose of smooth splicing, and finally the infrared digital holographic reconstructed amplitude and phase images are obtained.

2. The high-resolution infrared digital holography method based on convolutional neural network according to claim 1, characterized in that: A hologram containing multiple samples and different depth information is dynamically reconstructed to obtain a dynamically reconstructed hologram based on a convolutional neural network model; the loss function used in step three is an adjustable function.

Images