Applying 3D reconstruction technology to the field of medical plastic surgery is an effective means of assisting medicine to provide doctors with preoperative preparation data for reference. The current 3D scanners have high cost, short lifespan, large structure and difficult maintenance, which are not conducive to the popularization of this technology in the field of medical plastic surgery. A 3D imaging method of local human body features based on smartphone photogrammetry is proposed. The speckle pattern is projected onto the surface of the object to be measured by combining the micro projector. The modulated speckle pattern collected by the smartphone is combined with the prepared speckle template pattern. Between them, the local grayscale correlation of the image is used for matching to realize the reconstruction of the 3D model, and the complete 3D model is restored through the multi-view point cloud stitching technology. The efficiency and convenience of reconstruction are improved, and the utility model has the characteristics of low cost and simple operation. Taking the local features of the mannequin as the surface to be reconstructed, the 3D reconstruction of the continuous surface with weak texture is realized. The experimental results show that the average error of 3D surface reconstruction with this method is 0.43mm and the standard deviation is 0.30mm, which can meet the needs of actual measurement.

In this experiment, a dual-color passive synchronous all-polarization-maintaining fiber laser is demonstrated, which can generate wide spectrum for coherent anti-Stokes Raman scattering detection. The laser is composed of Yb-doped and Er-doped fiber lasers. Stable passive synchronization state can be achieved within 150μm cavity length mismatch distance by injecting synchronization. The spectrum of Er-doped laser can be broadened to 1000~1700nm by a highly nonlinear fiber, Then the Er-doped laser is temporally and spatially combined with Yb-doped laser to generate a sum-frequency signal through nonlinear crystal. The passive synchronous dual-color fiber laser can be used as a stable and ultra-fast light source for the detection of coherent anti-Stokes Raman scattering and is expected to realize the simultaneous detection of multiple Raman peaks in the wide spectrum.

The cylindrical lens type glasses-free 3D display system has developed rapidly in various fields in recent years. In order to further solve the problem of cylindrical lens type naked eye 3D crosstalk, a method is proposed to reduce the crosstalk between adjacent viewpoints of cylindrical lens grating autostereoscopic displays by shifting the pixel position on the LCD image panel, reducing the pixel interval and adjusting the tilt angle of the cylindrical lens. The causes of crosstalk for the dual-view cylindrical lens type naked eye 3D display system were analyzed.Paramefers such as the thickness,pitch,focal length and carvature radius of cylindrical lens were calculated. The relationship between pixel position, cylindrical Lens tilt angle,pixel interval, crosstalk area and moive thickness of LCD image panel is examined. The relationship between the thickness of the moiré fringe reduces the maximum crosstalk at the edge of the viewing zone while ensuring the thickness of the moiré fringe remains unchanged. The above method is simulated by Lighttools software, and the local area with the number of pixels of 120×60 is simulated. The results show that the maximum crosstalk at the edge of the viewing zone before and after the improvement is reduced from 50% to 30%, and the crosstalk at the best view point is reduced from 0.93% to 0.32%, and the best view point area is extended by ±15mm, which is of reference significance for reducing cylindrical lens-type naked eye 3D crosstalk.

Living cell observation reveals the life state of cells, which is the basis of many researches in biomedicine. However, most living cells are difficult to observe because of the transparency. Zernike Phase Contrast (ZPC) and Differential Interference Contrast (DIC) are usually used to enhance the image contrast. But the commonly used ZPC and DIC microscope are complex and expensive, and rely on Wollaston prism, phase contrast ring and other specific optical components. A multimodal phase imaging method based on dual-view transfer of intensity equation is proposed. With the defocus images acquired by two cameras installed in binocular tube of the microscope, the quantitative phase of the object can be obtained by solving the transfer of intensity equation, the ZPC image and the DIC image of the object could subsequently be computed. The results of numerical simulations and imaging experiments show that the dual-view multimodal phase imaging method can provides high quality ZPC and DIC image. The proposed method doesn’t rely on specific optical components or complex configuration, and is low-cost and easy to implement. Thus, it can substitute the expensive ZPC microscope and DIC microscope in the application of unlabeled biological imaging.

In order to eliminate the parasitic fringe error introduced by PBS adhesive layer in michelon’s measurement optical path, a dynamic interference optical path with TP (twin PBS) structure as the light splitting core is proposed. Starting from the Mach Zehnder optical path, the structure integrates the polarization optical requirements of synchronous phase-shifting, the requirements of open external measuring arm, the requirements of compactness and adjustment convenience, and the requirements that the test beam and reference beam are separated only on the coating surface in PBS. The polarized light transmission model is constructed by Jones matrix. Through the simulation of ZEMAX and other software, the compact design and adjustment process flow of the new optical path are designed, and the adjustment steps are compressed. Finally, the TP dynamic interferometer composed of light source module, imaging module, modular optical path and polarization CCD is used to pre measure the three mirror system before coating, and the precision measurement under (4%)3 extremely weak reflected light is realized.

Wavelength beam splitter is one of core components in the field of optical communication. In order to meet the design requirements of increasingly complex micro-nano devices, a new intelligent inverse design method which combining the Finite Element Method (FEM) with the Method of Moving Asymptotes (MMA) and the Material Interpolation Method (MIM) has been presented, which has a faster calculating speed and a lower arithmetic complexity, and adapt to the micro-nano photon devices design. This algorithm can complete the optimal design according to the objective function given by the designer, with low professional requirements for designers and wide range of adaptations. Based on this algorithm, a three-channel chip-integrated wavelength beam splitter with footprint of 2μm×2μm is successfully designed. Its transmission rates of the three channels are 47.5%, 53.4% and 52.5%, respectively, and the extinction ratios are 36, 9.6 and 20. Which provides a new idea and reference for design of micro-nano devices.

The principle of external sphere grinding of hemispherical Resonator based on generation method is discussed. The theoretical calculation model of curvature radius was established. The dominant errors affecting the machining precision of external sphere are shape error of diamond tool, positioning accuracy of CNC machine axis and runout of CNC machine shaft. The shape error of diamond tool includes the diameter error of diamond tool and radius error of diamond tool edge. The positioning accuracy of CNC machine is determined by the positioning precision of X-axis, Y-axis and B-axis.The runout of CNC machine shaft includes axial and radial runout errors. After optimizing the machining process parameters, the roundness of the hemispherical resonator is less than 0.25μm, the concentricity of the hemispherical resonator is less than 0.4μm, the surface roughness is less than 0.02μm and the Q value is greater than 2.6×107.

In view of the low detection accuracy caused by color bias in traditional face detection, the deep learning method realizes face detection by training a large amount of data, resulting in high hardware requirements. A simple convolutional neural network is proposed for face and non-face recognition, and then white balance algorithm is used to solve the problem of color cast. Skin-Color detection was realized by combining YCgCr color space with K-means clustering. Finally, face detection is realized on the basis of skin color detection. Its accuracy is about 3% higher than the traditional face detection method, and its speed is about twice faster than the face detection based on deep learning.

Multi-channel narrowband filters have important applications in the field of multispectral imaging. There are many important parameters of filter optical performance evaluation, such as center wavelength of each channel, half-height full width of passband and cut-off band. It is significant to obtain the influence of spectral characteristics of narrowband filter on coating process and multispectral imaging application by accurate testing. In this work, the difficulty of spectral crosstalk in the testing process of multi-channel narrowband filters was analyzesd. A testing method, principles and equipment of measuring the spectral characteristics of multi-channel narrow-band filters based on microscopes and fiber spectrometers were demonstrated. The influence of spectral resolution on the test results was analyzed through simulation experiments. Additionally, the original data was corrected by the five-point correction method of differential operator, and the errors were also analyzed. The problems of spectral crosstalk and spectral resolution in the process of multi-channel filter testing were solved, and the results can be used to guide the production of filters and provide accurate spectral information for spectral imaging.

In interferometry, environmental, imaging system thermal noise and ununiform reflectance of the test surface all lead to inconsistent phase distribution, which in the form of residues occurred in the wrapped phase maps. Under this condition, how to completely recover the phase unwrapping result with high-precision becomes to be an unavoidable issue. Different from conventional unwrapping techniques, which construct branch cuts between the residues to separate the noise-affected areas; or different from the approaches implementing minimize norm methods, at the cost of losing local information of the details, to the best of my knowledge, harmonic mean, is for the first time applied in phase unwrapping. Based on its specific property of handling large outliers, the four neighborhood pixels around the residue are analogous to the harmonized mean data, giving rise to a straightforward way to identify the noise-related pixels and to obtain a global unwrapping result. Finally, a computer simulated complex surface with high level of noise is used to verify the accuracy, efficiency and simplicity of the proposed algorithm in implementation, through both quantitative and qualitative analysis.

Pulse wave is an important indicator of human physiological status. The non-contact pulse wave detection technology based on imaging has vital research significance in the field of medical and health. Because the current non-contact method of obtaining pulse wave is always lack of waveform details, a method of obtaining pulse wave signal using near-infrared light source and multi-frame continuous photography of carotid pulse was proposed. With experiment environment of an 850 nm light source, epidermal vibration caused by carotid artery pulsation was photographed using an industrial near-infrared camera. The original pulse wave signal was extracted after selecting the region of interest, then the signal was filtered by maximal overlapping discrete wavelet transform. Eventually, the preserving details pulse wave was obtained. The measurement results of different subjects showed that the obtained wave observes the characteristics of the main wave, tidal wave and heavy pulse wave. The proposed method plays an important role for non-contact acquisition of pulse wave.

In order to test the ablating accuracy and stability of excimer laser corneal ametropia cure system in refractive surgery, a method based on depth measurement of PMMA plate is proposed. Firstly, the theoretical value of corneal ablation depth is analyzed. Then the reasons of the difference between the measured ablation depth and estimated ablation depth of different excimer laser machines are analyzed. Finally, the relationship between the estimated ablation depth of cornea and the measured depth of PMMA plate is fitted and the fitting line is taken as the theoretical value. The experimental results show that by measuring the depth of PMMA plate pre-operation in operation mode and comparing it with the theoretical value, the error of the indicated value is within ±5%.

In order to solve the problem that it is difficult to distinguish the upper and lower surfaces in the surface defect detection of transparent media, a single-sided imaging method of glass surface based on Fresnel formula is proposed. In this method, the change law of light polarization characteristics after reflection and refraction of linearly polarized light is used to optimize the incident angle and polarization direction of incident light, so that the polarization direction of imaging light on the upper and lower surfaces is close to orthogonal, and then the filtering characteristics of polarizer are used to select the required light for imaging. The optimal incident angle and polarization direction are obtained through theoretical calculation and simulation analysis, which are verified by experiments on the spectrometer, and the evaluation function is set to evaluate the imaging at different incident angles. Experiments show that this method can effectively eliminate the reflected light on any side of the upper and lower surfaces of transparent media, so as to realize single-sided selective imaging. After the evaluation and analysis of the evaluation function, it is found that the imaging effect is better when the incident angle is small. The scheme is simple and reliable, will not damage the test samples, and is easy to put into actual production.

In non-contact 3D optical measurement based on deformation fringe pattern analysis, phase distribution is extracted from the collected deformation fringe pattern to obtain the surface information of the measured shape. But the acquired fringe pattern contains noise in measurement, which affects the accuracy of extracting phase information. In order to remove the noise in fringe pattern better and faster, an improved U-net neural networks filtering algorithm based on deep learning is proposed. In the field of image denoising, U-net acquires few shallow features. The proposed method contains 1×1 parallel convolutional branches in the convolutional of U-net, which is used to obtain multi-scale feature. And adds 1,2, 3 1×1 parallel convolutional branches for experiment. Fringe pattern with high-density regions is used in the experiment, and the proposed method is compared with the state-of-the-art deep-learning fringe pattern denoising algorithm. The denoising effect of the proposed method is improved by 0.9%, the denoising efficiency is improved by 41.7% and the training time is reduced by 30.8%.

Aiming at the situation of occlusion, light change and complex scenes in current gesture recognition, which easily leads to the decline of the accuracy of gesture recognition, a gesture recognition scheme based on the coupled depth camera of somatosensory control is proposed. In order to improve the accuracy of gesture recognition, the proposed gesture recognition scheme includes Leap Motion data and Kinect depth data. Firstly, the distance between the fingertip and the center of mass of the hand, the height between the fingertip and the palm plane, the angle between the fingertip and the center of the palm, and the 3D position of the fingertip in the hand reference system are extracted from the somatosensory control data. Then, the distance between the finger sample and the center of the hand, the local curvature of the hand contour, the similarity between the distance features and the connected area of the hand shape are extracted from the Kinect depth data. Then, in order to combine the complementary information of two kinds of sensor data and discard redundancy, a joint calibration method is defined by finding the rotation and translation parameters through the 3D position of fingertips collected, and minimizing the average re-projection error of all fingertips in all acquisition frames, determine the external parameters of the somatosensory control sensor and the Kinect depth sensor, and complete the coordinate conversion of the two sensors. Finally, a classification learning method based on multi-class Support Vector Machine (SVM) is proposed. Experiments show that the average recognition rate of Jochen Triesch gesture database is 97%. In different light, skin color and background environments, The proposed algorithm has higher accuracy and robustness than the existing hand gesture recognition algorithms.

Automatic segmentation of pancreas has always been a challenging problem in medical image segmentation. The pancreas is an organ with a high degree of anatomical variability, and it is difficult for the current multi-atlas segmentation methods to accurately segment the edges of the pancreas. Focusing on this problem, a segmentation algorithm is adopted based on multi-atlas registration to segment the pancreas, and optimizes a post-processing method of local dynamic threshold. In the label fusion stage, four label fusion algorithms are used for comparison: probability threshold fusion algorithm, Majority voting (MV) algorithm, STAPLE algorithm and SIMPLE algorithm. In the post-processing stage, the local dynamic threshold processing method is adopted. First, the target area is extracted from the target image through the preliminary segmentation result, and then the threshold value is automatically determined to realize the binarization of the area. Finally, the intersection with the preliminary segmentation result is taken as the final segmentation result. A leave-one-out cross-validation strategy was used to segment 80 NIH pancreatic CT images and 22 pancreatic CT images from local hospital at Shanghai, and the final DSC obtained were 79.98% and 81.30%, respectively. The experimental results show that the proposed method achieves effective segmentation of the pancreas.

Aiming at the problem that the current image super-resolution reconstruction method fails to make full use of the global and local information of the image, which causes the reconstruction result to lose part of the source image information to a certain extent, a multi-scale dense residual network is proposed to achieve this. Super-resolution reconstruction of the image. The network is based on dense residuals and integrates the multi-scale feature information of the image, ensuring that the network does not lose feature information in depth while obtaining more information under different receptive fields, thereby avoiding excessive loss of information in the original image. In addition, in order to recover high-resolution images containing enough high-frequency information from low-resolution images with low-frequency redundant information, the network combines spatial attention and channel attention to process low-resolution features at different scales in an unequal manner. This method can effectively highlight the high-frequency components in the feature map, so that the network can better learn and fit the feature information of the label image, and restore an image that is close to the real image. A large number of experimental results prove the effectiveness of this method.

Optical coherence tomography (OCT) is widely used in ophthalmology to observe the morphology of the retina, and is of great significance for the detection and diagnostic evaluation of lesions. Due to series of retinal diseases caused by liquid accumulation, a neural network with global context feature information is designed for liquid detection and segmentation in retinal OCT images. By means of multi-scale feature extraction and fusion,a multi-scale parallel extraction and highly integrated U network model PH-UNet is proposed, which is a new deep convolutional network for liquid area segmentation in OCT images. PH-UNet network captures multi-scale contextual information, and better utilizes information extraction and fusion methods to perform end-to-end segmentation of the liquid area of the OCT image. The proposed model is segmented on three types of retinal fluid. The intraretinal fluid (irf), subretinal fluid (srf), and pigment epithelial detachment (ped) are segmented and compared with other classic segmentation network models. The best results have been achieved on the three indexes of precision,dice and mIoU, which proves its superiority.

In order to improve the detection accuracy of Visible-Infrared Cross-modal Person re-identification, a dual flow network model based on Visible-Infrared Cross-modal bidirectional feature generation is proposed. Compared with the existing algorithms, employ the bidirectional feature generation method to transfer the Cross-modal pedestrian features, which significantly enhances the Cross-modal feature expression. At the same time, the dual flow network is used to extract the discriminative dual-modal features, and the coarse-grained and fine-grained loss fusion strategy designed improves the accuracy of cross-modal pedestrian retrieval.The experimental results indicate that compared with the latest method, the average accuracy of cross modal pedestrian recognition, 92.91% on RegDB and 66.17% on SYSU-MM01, is increased effectively by proposed method.

The classical RX anomaly detection operator assumes that the background data information conforms to Gaussian distribution, but the hyperspectral image is degraded due to a large amount of additive noise, and the background information does not conform to this kind of distribution. To solve this problem, the algorithm of RX anomaly target detection for hyperspectral imagery based on low-rank tensor decomposition is proposed. Firstly, the low rank tensor decomposition method is introduced to recover the hyperspectral image, and which uses the tensor data structure and low rank data characteristics of hyperspectral image, so that the anomaly target information becomes prominent compared with the complex background information, and then the RX anomaly detection operator is used to detect the anomaly target in the recovered hyperspectral image; Finally, the anomaly target detection results are obtained. Through the comparison of simulation experiments, the new anomaly target detection method has the characteristics of high detection accuracy, low false alarm rate and good robustness.