Phase unwrapping occupies an important position in digital holographic microscopy imaging technology. It is an indispensable key step for obtaining phase information. The current traditional phase unwrapping algorithm has entered the platform stage. Deep learning for phase unwrapping research methods were used and better experimental results were obtained. This study proposes an improved U-Net phase unwrapping method, adding the channel attention module after the residual block, and uses deep separable convolution to replace part of the traditional convolution. This study uses a large number of analog phase maps generated by random matrix as data training sets to achieve the purpose of phase unwrapping, and has solved the problem that optical path is inconvenient to obtain a large number of data sets. The proposed method is used to unpack different experimental holograms, and compared with the results of other unpacking algorithms, the experimental results show that the edge of the proposed method is smoother and the background is flatter. The structural similarity index of the experimental test set and the Discrete Cosine Transform unwrapping results increased from an average of 0.932 to 0.973, and the peak signal-to-noise ratio increased from an average of 21.60 to 29.18.

The zero-frequency component (ZFC) distributes widely and carries no phase information of the object in off-axis hologram, which significantly reduces the signal-to-noise ratio and thus decreases the accuracy of the reconstruction result. To suppress ZFC, existing methods need record multiple holograms or suffer from manually adjustment of hologram decomposition layer to achieve optimal results. Laplace operator is also used to suppress ZFC but it is difficult to extract the +1 level. To overcome these problems, a Hilbert-Huang transform based method is proposed that only needs one hologram and avoids manually adjusting the layer of decomposition. To verify the effectiveness of the proposed method, the frequency spectrum of the hologram processed by the proposed method is compared with the frequency spectrum of the hologram without ZFC suppression, the hologram suppressed with Haar wavelet transform and Laplace operator. Three regions in the corrected phase maps of the holograms are randomly selected to obtain the standard deviations and sectional diagram. The experimental results show that the proposed method can effectively suppress the ZFC and performs better than other two methods.

With around 70% of the earth's oceans and seas, there is a huge amount of resources beneath the surface to be explored, exploited and utilised. However, due to the complexity and variability of the underwater environment, it is not easy to explore and discover the underwater world. The introduction of polarimetric imaging, with its multi-dimensional polarisation vector information compared to traditional optical imaging techniques, has attracted much academic attention and has led to a boom in the development of underwater polarimetric detection and processing techniques. When imaging underwater targets, the large number of particles and some soluble materials in the water column can cause absorption and scattering of light, resulting in severe attenuation of light energy entering the receiver, resulting in blurred and poor quality underwater images that are difficult to achieve for practical applications. The polarization characteristics of underwater targets under different illumination are the research content of this paper. CMOS cameras are used to polarize objects of different materials under different light sources. The polarisation information is fused to combine the advantages of the different polarisation vector images, and finally the image is enhanced using the Laplace edge enhancement technique to recover the degraded image.

In order to solve the challenge of freeform surface measurement, a large-range laser differential confocal sensor with high precision was constructed. Based on the principle of laser differential confocal focusing, the sensor uses normalized single-point focusing model to eliminate the influences of the surface reflectivity and tilt of the capsule. A non-equidistant fractional-order accumulation model is used to predict the focus position and achieve fast tracking and focusing. The aerostatic linear guideway and PZT are combined to expand the axial scanning range. The experiment results show that the axial resolution of the sensor is better than 3nm and the measured surface tilt angle can reach 22°. The axial scanning range can reach 50mm and the zero-crossing standard deviation of the differential curve is 5.7nm, which has good repeatability. The sensor provides a new method for measuring the profile of freeform surface.

The accuracy of mode field analysis directly affects the performance of devices based on liquid crystal optical waveguides (LCOWs). The following procedures were used to accurately solve the mode field distribution in LCOW, with a dielectric tensor being graded along one dimension. The anisotropic magnetic field coupling equations of liquid crystal are discretized by the finite-difference method, under the case that the gradient characteristic of liquid crystal field-induced reorientation has been considered. Then, the eigenvalue equation that can accurately solve the modes in the LCOW is obtained by improving the traditional difference operator matrix. Finally, a numerical example is given to compare the solution results with the coupling relationship between the field components in the eigenmode, which is to verify the effectiveness of the proposed mode solver.

A miniature industrial endoscope system based on biprism and single camera is proposed, which can obtain two views of the same scene at the same time, and display on single-frame CCD image. By establishing the system structure parameter model, the influence of the measurement range on the system measurement error is analyzed. The result shows that, in the close measurement the system measurement range is small and the error is small. As the measurement range becomes larger, the error will also become larger, and the error changes gently along the straight line ω1=ω2. Using this system to measure the grid size of the high-precision calibration board, the measurement error reaches the order of microns, and the average error is 2.06%, which meets the needs of industrial applications. The system has simple operation, compact structure and relatively few parameters. It has certain guiding significance for the structural design and application of industrial endoscope.

A Yb-doped picosecond fiber laser with 35.4W for average power, 12ps for pulse duration, and 300kHz~1MHz for repetition rate is designed based on cascade amplification. The oscillator and the pre-amplification system are comprized of all fiber. Rod-type photonic crystal fiber is used as the gain medium of the main amplifier. The output repetition rate can be adjust from 300kHz to 1MHz by an Acousto-Optic Modulator (AOM). The all-fiber design of the oscillator and the pre-amplifiers has the characteristics of simple optical path structure and high environmental stability, which is expected to promote the industrialization process of high-energy femtosecond fiber laser and improve its adaptability under complex environment.

With the development of ultrafast laser cold micromachining, ultrafast laser has been gradually investigated and applied on the protection of cultural relics. According to the application requirements of cleaning the surface pollutants of the ancient city wall cultural relics, the marble cultural relics and surface pollutant simulation samples of the city wall are made, and the experimental research on the cleaning of cultural relics samples based on picosecond laser is carried out, and the effect of nanosecond laser cleaning is compared. Confocal microscope observation and fluorescence spectrometer composition analysis were used in the experiment. Change the laser cleaning parameters such as laser output energy, cleaning times and scanning speed, on the basis of not damaging the marble substrate and comprehensively considering the cleaning efficiency and cleaning effect, to optimize the experimental parameters. The experiment found that when the laser parameters are power 18W, 1000mm/s scanning speed, 8 times. After cleaning, the proportion of sulfur element decreased by 94.57%, and the surface roughness of the area after cleaning was 1.267μm. According to the proportion of pollutants after cleaning, the surface roughness, etc., the overall effect of picosecond laser cleaning marble cultural relics samples is better than that of using nanosecond laser.

Radiation monitors using Al2O3∶C as sensitive material have advantages of small size, high sensitivity, online remote annealing capability, etc. Accurately measuring OSL spectrum of Al2O3∶C not only helps to better understand the energy level distribution and luminescence mechanism of Al2O3∶C, but also guide the design of luminescence collecting and measuring system. Two different methods, direct measurement and dichroic measurement, are used to accurately measure the OSL spectrum of Al2O3∶C, and the results consistent with each other. The obtained spectrum has a peak shape, centered at ~414nm and with a full width at half maximum of ~62nm, and the long-wavelength side decreases more slowly. The broadband luminescence centered at 414nm (corresponding to energy of 3.0eV) is attributed to radiative transition from the 3P excited state to the 1S ground state of F-center. The dichroic spectrum measurement method designed can completely eliminate the influence of high-density reflected stimulating light, which can also be used in measurement of spectrum of other OSL materials.

The phase-shifting interferometry technique is widely used in wavefront detection interferometers due to its high measurement accuracy. The phase shift error is the main source of error in measurement process. Based on a self-tuning phase-shifting interference algorithm, the accuracy of wavefront phase restoration is studied under calibration errors and random phase shift errors. For the calibration error, the algorithm can accurately calculate the actual phase shift step size,thus greatly improving the phase restoration accuracy.Compared with the classic five-step Hariharan algorithm, the simulation results show that the phase restoration PV (Peak-Valley) and RMS (Root Mean Square) error response of this algorithm is lower, and its PV error response is much lower than 10-3λ, where λ is the center wavelength of the light source, while the Hariharan algorithm is on the order of 10-3λ. Based on the phase solving process of the self-tuning algorithm in phase shift calibration error, the algorithm is extended to be more suitable for random phase shift error. Within the same 20% random phase shift error range, the absolute value of the deviation from the Hariharan algorithm calculation result is close to 10-9λ, which can achieve high restoration accuracy. The self-tuning algorithm is used in the measurement of the surface topography by the interferometer. The experimental results show that compared with the Hariharan algorithm, the self-tuning algorithm can obviously suppress the ripple error when there is only calibration error. There is a deviation in the surface PV. In a small vibration environment, the phase restored by the two algorithms are highly consistent.

In order to reduce the manufacturing cost of optical gas concentration sensor, a meter-optical-path infrared gas absorption cell structure is designed and optimized by using ZEMAX optical design software. The structure consists of two parabolic mirrors for collimating and focusing light and five planar mirrors for increasing optical path. The simulation results show that the geometric loss of the gas absorption cell is only 1%, and the optical path is 1049.75mm. Compared with the gas absorption cell with traditional lens structure, the structure has the characteristics of small volume, long optical path, low loss rate and strong expansibility, and can be used to construct low cost and sub-ppm level infrared gas sensor.

Wide-angle lenses are generally used in photoelectric measurement systems with large fields of view, which will cause serious distortion problems in the obtained measurement images. In order to accurately calibrate this type of large-distortion camera, the radial distortion division model and the corner sub-pixel coordinate extraction method are used to first solve the image distortion center coordinates and distortion parameters. and then uses the two-dimensional plane checkerboard marker points and The corresponding relationship of the image points is used to solve the homography matrix, and then the internal and external parameters of the camera are further solved according to the homography matrix. After solving the relevant parameters, use the Levenberg-Marquardt method to iteratively optimize the solved parameters, and then based on the iterative optimization, the reprojection error data is eliminated according to the 3σ rule. Calibration is performed until all data meet the requirements, and finally high-precision correction of large-field-of-view distorted images can be achieved. In order to verify the effectiveness of the proposed method, calibration experiments on simulated images and actual images are carried out . The results show that this method can appropriately improve the calibration accuracy. In the actual experiment, the average mean square reprojection error is reduced by 0.0103 pixels, which is equivalent to improving the correction error accuracy by 0.7%.

In order to realize the high precision calibration of the Los Direction of the RC system camera with small aperture, the establishment of the calibration process coordinate system and the conversion method are discussed, and a new calibration method is provided. The cross-wire is imaged on the CCD image plane by using the theodolite to collimate the beam, the image and the background are separated by using the method of maximum variance between classes, and the centroid of cross-image is extracted by Harris Corner Detection Algorithm, the high precision calibration of the Los of the RC system camera with small aperture is realized. The result shows that the calibration accuracy reaches 1.7″, which is 12 times higher than the traditional theodolite calibration accuracy of 20.9″. It provides a new accurate and feasible method for camera sight axis calibration, and can be applied to other camera sight axis calibration.

The use of Optical Coherence Tomography (OCT) to produce retinal disease images is an important measure to classify ophthalmic diseases. The purpose of this study is to use the transfer learning method of four different classification models to automatically classify the OCT retinal images of diabetic macular edema, age-related macular degeneration, and drusen to realize the application of transfer learning in OCT image classification. After pre-training the four neural network models VGG-16, Inception V3, MobileNet-V2, ShuffleNet-V2 on the large-scale graph classification data set, fine-tunieg the model and update the training parameters to find the realization of the above three ophthalmic diseases. The optimal model of automatic classification achieves efficient OCT retinopathy classification effect. The experimental results show that the lightweight MobileNet-V2 of the four models has better evaluation indicators than other models after the model is fine-tuned.

3D glioma magnetic resonance imaging has different tumor shapes and blurred edges. The segmentation method based on 2D Convolutional Neural Network cannot segment the three-dimensional image well. In order to accurately segment the tumor in the three-dimensional image, a 3D Convolutional Neural Network brain tumor image segmentation method fused with multi-scale feature information is proposed. The feature information is extracted by parallel 3D dilated convolution, and the information of different receptive fields is fused. The Dice loss and the BCE loss are combined to form a new loss function and cooperate with the identity mapping to further improve the segmentation accuracy. The model was verified on the BraTs2020 data set. The Dice coefficients of the whole tumor area, core area, and enhancement area segmented by the model are 89.1%, 83.9%, 82.6%. The model was verified on the LGG brain tumor image data set, and the Dice coefficient reached 93.3%. The segmentation method can not only accurately segment three-dimensional glioma images, but is also suitable for segmentation of two-dimensional glioma images.

In order to improve the stereo matching accuracy, a fusion stereo matching algorithm of adaptive SAD with super-pixel segmentation constraints and Census algorithm is proposed. To address the errors introduced by the indiscriminate use of the grayscale values of the pixel points within the window in the stereo matching process of SAD, First the super-pixel segmentation method of Simple Linear Iterative Clustering (SLIC) is used to process the map to be matched, and the segmentation results are combined with the distance between the neighboring pixel points and the center pixel point within the window to assign appropriate weights to the grayscale values of the pixel points within the window in the SAD stereo matching process; Subsequently, the Census stereo matching process is performed, and the matching results of the two algorithms are adaptively fused; Finally, post-processing processes such as left-right consistency detection and occlusion point filling are performed on the Initial parallax map. The experiments show that the algorithm is significantly better than the traditional algorithm in terms of matching effect, can be well adapted to detail-rich images and has better adaptability to groups with vertical shifts, and is robust to image contrast and illumination changes.

In order to further improve the brightness and contrast of the infrared image while maintaining the natural effect of the image, an infrared image enhancement method based on adaptive Gamma correction is proposed. The improved bilateral filter is used as the central surround function of Retinex to extract the low-frequency and high-frequency parts of the infrared image, that is, the base layer and the detail layer. Gamma correction adaptive to the proportion of pixels in the dark area is performed on the base layer image, so as to improve the brightness and definition of the image, and adaptive piecewise linear stretching is applied to the detail layer image to further improve the contrast of the image. Experimental results confirm the effectiveness of this method, compared with the existing methods, the visual effect of this method is clearer, and the detail information is richer.

Concerning the poor fusion quality of the infrared and visible light outdoors image, an infrared and visible light outdoors image fusion method based on convolutional neural network is proposed. Firstly, this method makes use of rolling guidance filter to preprocess the input infrared image, so that the noise is filtered and the useless information is eliminated. Then, the Curvelet transformation is used to decompose the infrared image and visible light image, the images are decomposed to high frequency coefficients and low frequency coefficients, the deep feature fusion rule based on convolutional neural networks is utilized to fuse high frequency coefficients, and the minimum fusion rule is utilized to fuse low frequency coefficients. Experimental results show that the images fused by the proposed method achieve better subjective visual quality and objective quantitative result.

Infrared night vision technique is effective on the security enhancement of sea early warning detection system, but the sea images captured by the infrared night vision viewer usually contain an amount of “background clutter”, it leads to obvious detection performance reduction of sea micro targets. To solve this problem, both active contour model and interpolation filter are combined to propose an accuracy micro target detection algorithm of infrared night vision. Firstly, this algorithm uses global active contour model and local active contour model to search the region of interest, respectively, thus by use of double-layer active contour models, the bad impact of noise and clutter on target detection are eliminated. Then, the interpolation filter with direction variation is proposed to filter the region of interest along the edge, in order to reduce the interference of week edges with the target real edge. Experiments and analysis are carried on the real infrared night vision sea image set, the results show that the proposed method can improve the detection performance of sea micro targets, this algorithm is good for sail on the sea.

To resolve the problem of the high false alarm rate of traditional adaptive target detection methods of infrared night vision, an adaptive small target detection method of infrared night vision based on sliding window is proposed. First of all, the Harris hawk optimization algorithm is enhanced with combination of teaching learning based optimization algorithm, in order to enhance its ability of escaping from local optimal, a mixed multiple population Harris hawk optimization algorithm is designed; Secondly, a fitness function of small target based on sliding window is designed; Finally, the infrared night vision small target is detected adaptively. Experimental results based on public datasets suggest that, compared to the other adaptive infrared small target detection method, the proposed method has lower false alarm rate for small target detection.