As a new nucleic acid detection technology, dPCR has been widely used in biomedicine and other fields. In order to further improve the detection efficiency of the existing dPCR detection system, a set of dPCR detection system based on large field of view fluorescence microscope is designed. The illumination system and filter module are designed in the microscope object space, and combines the compound eye illumination system to achieve a large area uniform illumination, the illumination area can reach 32mm×22mm, and the illumination uniformity can reach more than 80%. A method for counting and detecting petri dishes combined with neural network is proposed, which can effectively pick out the petri dishes that have evaporated in the petri dishes. The system only needs a few seconds from the camera exposure to the end of detection. Compared with traditional image stitching detection methods, the detection efficiency is improved significantly.

After the laser beam with a center wavelength of 1064nm passes through the collimated beam expander optical system, its light intensity still presents a Gauss-like distribution. In order to carry out the research on shaping and homogenization, a gradient attenuator was designed according to the aperture method and the laser light intensity distribution curve. The structure of the attenuation sheet is composed of a plurality of light-shielding rings and light-transmitting rings, which physically shield the laser beam to achieve the purpose of reducing the inhomogeneity of the laser energy. MATLAB is used to calculate the radius of each light-shielding ring zone, and Zemax is used to simulate the incoherent irradiance results of the attenuator when it acts on the laser. The simulation results show that the attenuator can reduce the inhomogeneity of laser energy. Adding the attenuator to the improved beam expansion system, the measured laser energy inhomogeneity is 9.31%. The measurement results show that the concentric ring-band gradient attenuator designed based on the diaphragm method is effective in reducing the inhomogeneity of laser energy and has certain engineering reference significance.

An effective assembly method is presented for 1.3m aperture coaxial quad-inverse lens. The primary mirror has a diameter of 1.3m and is supported by a back bipod. The precise positioning of the supporting structure is carried out by using the laser tracker polygon measurement method, and the position of the primary mirror are adjusted by using the Stewart mechanism position inverse solution method. Surface measurement is made by changing the orientation of the reflector assembly to extract the surface error of the reflector caused by gravity. The optical datum of the reflector and the lens datum can be transferred by centering the reflector based on interferometry using the Offner zero-position compensation detecting light path. The lens is mounted vertically, and the assembly accuracy and convergence speed are improved by testing the lens image height and calibrating the standard mirror shape online. The center and edge field wavefronts RMS of the lens are 0.053λ (λ=0.6328um) and 0.077λ, respectively.

The solar simulator can accurately simulate the collimation, uniformity and spectral characteristics of solar irradiation, it has a high simulation precision of heat flow outside space. It is mainly used for spacecraft thermal balance test, thermal coating property test and material aging test, which can effectively test the satellite's light irradiation performance and verify the spacecraft's thermal design. An integrator component is designed according to the characteristics of large and medium-sized solar simulators, which can withstand high temperature load and ensure the uniformity and irradiation energy of large and medium-sized solar simulators. The principle of integrator module of solar simulator is introduced, and the structure design, thermal design and simulation calculation of integrator module are carried out by LightTools and Flunet. The design results show that the irradiating surface of medium-sized solar simulators can reach Ф2000mm and the nonuniformity can reach ±4%. The experimental structure shows that the irradiating surface of medium-sized solar simulators can reach Ф2000mm, and the nonuniformity can reach ±3.75%, the maximum irradiance can reach 1.39 solar constants, which can meet the needs of large and medium-sized solar simulator.

When testing tracking accuracy of the optoelectronic tracking terminal by optical dynamic target, in view of the problem that surface-shape error of the rotating photoelectric collimator caused by the centrifugal field is too large, a method based on flexible beam is proposed to reduce the surface-shape error of primary mirror of the rotating collimator under centrifugal condition. Firstly, the optimal radial position of the supporting point of the primary mirror is solved; Secondly, the key design parameters of the flexible supported beam on the back of the primary mirror are optimized on the Isight optimization platform; Finally, Sigfit software is adopted to analyze the surface-shape error under centrifugal condition. After adopting the optimized scheme of the backside support, under the influence of 1.64g centrifugal force, the surface accuracy RMS of primary mirror is better than 0.03λ (λ= 632.8nm), and the first-order natural module frequency of the collimator is 167Hz. The 4D interferometer is adopted to detect the wave aberration of the collimator, the result indicated that?the RMS value of wavefront aberration of collimator is better than 0.067λ. Through simulation analysis and testing experiments, it is proved that the dynamic stiffness and surface shape accuracy of the rotating collimator with flexible support meet the application requirements of optical dynamic targets well.

The two-layer and three-layer antireflection coatings composed of silica and silicon nitride were designed and optimized by finite-difference time-domain (FDTD) method. High quality silicon dioxide and silicon nitride films were deposited by plasma enhanced chemical vapor deposition (PECVD) to prepare silicon oxide and silicon nitride double-layer antireflection coatings. At the same time, three layers of antireflection films, including silicon nitride, silicon oxide and silicon nitride, were prepared. The antireflection effects of two kinds of antireflection coatings were measured. The results show that the reflectivity of double-layer antireflection coatings is less than 0.18%, and the three-layer antireflection coating has larger bandwidth.

Based on the Z/X/B triaxial CNC machine, a constant tangent-face angle(CTFA) method for high-order aspherical grinding is proposed. In this method, equal chord length is used as the criterion to dissociate the profile of the aspherical, and the tangent-face angle is assumed as a constant to calculate the running process of CNC machine. The equal chord length criterion can avoid excessively step of the profile dissociation which was caused by the slope gradual change, and CTFA method makes the grinding wheel to finish the grinding process kept using front face edge, which helps to achieve a aspherical surface in high precision and consistency. In addition, the grinding process with CTFA method is simulated by computer programming and the result shows that, parameters δ and C should make the running of CNC machine meets the conditions of no physical collision between the work piece and machine components, no reverse running of CNC machine in treble Z/X/B direction, running step must not smaller than the resolution of CNC machine and to make the process route of the machine as short as possible. The deviation of aspherical surface caused by grinding wheel abrasion can be relieved by reprogramming the CNC process after the working edge was sharpened or take the abrasion edge of the wheel as the working edge.

Changes of blood vessel morphology are closely related to disease. Diameter is the main parameter of blood vessel morphology and measurement of blood vessel diameter is beneficial to the screening and prevention of diseases. A method of measuring blood vessel diameter based on clustering algorithm was proposed to measure microvessels. Noise is present in most microvascular images (such as optical or photoacoustic microimaging) and the microimages can be enhanced by nonlinear transformation functions. Trained U-Net network model was used to achieve extraction of retinal vessels. The blood vessel diameter was measured by combining clustering algorithm and ray algorithm. Experiments show that this proposed algorithm was consistent with the traditional measurement results (P＞0.05). Compared with the traditional algorithm, the measurement accuracy of this algorithm was improved, and the measurement error was reduced from 4.21% to 2.27%, which meets the accuracy requirement of vascular measurement.

As a non-contact and three-dimensional mirror surface measurement method, stereo deflectometry has developed rapidly in recent years. In order to solve the problem that it is difficult to measure large inclination surface with traditional stereo deflectometry, the three-dimensional stitching of stereo deflectometry based on marker is proposed. In this method, the three-dimensional point cloud in the sub-regions under different perspectives was measured by stereo deflectometry. Meanwhile, the three-dimensional coordinates of the markers on the sample table were measured by binocular stereo vision. The markers coordinates were used for coarse stitching, and the ICP algorithm was used for fine stitching. A three-dimensional stitching measurement system is actually built, and a smooth convex spherical reflector with an inclination angle of 25.8 degrees is measured. After stitching, the spherical fitting error is within 3μm. The experimental results verify the feasibility of the method, which is of reference significance for the application of the stereo deflectometry to the three-dimensional measurement of complex optical surfaces with large aperture and large inclination.

Gear tooth profile deviation is one of the important parameters in gear verification. A new type of composite gear line structured light measurement system is used to measure the tooth profile of the first-level standard (u=1μm, k=3) gear involute template, and analyzes from the mechanical, thermal, optical, and electrical aspects. The source of measurement error, and based on JJF1059.1-2012 "Measuring Uncertainty Evaluation and Guidelines", evaluated the uncertainty of measuring gear tooth profile deviation with this system. Research shows that there are errors when using this measuring system to measure gear tooth profile, including: instrument measurement repeatability error, involute model standard error, model linear expansion error, center coaxiality error, gear installation eccentricity error, line structure Optical system error, etc. The combined standard uncertainty caused by CCD camera calibration error, structured light vision model calibration error and line structured light strip center detection error is 0.98μm, which is the main error of the new composite gear line structured light measurement system Source, and through analysis and calculation, the extended uncertainty of the measurement system is 2.3μm(k=2), which can meet the gear measurement with a measurement accuracy of 2.3μm or more.

In order to improve the recognition accuracy of traditional fall detection system and reduce the recognition time, a new fall detection model is proposed. The skeleton node obtained by depth vision sensor of Kinect V2 is used as the sample data source, and the improved k-means algorithm is used to calculate the clustering center point, and on this basis, the fall detection feature data is extracted. After reconstituting the feature data into 5×5 training sample data, the convolution neural network model is input to train and learn, and the optimal fall detection model is obtained. Experimental results show that the new detection model has higher recognition accuracy and faster operation speed than the traditional fall algorithm, which guarantees the real-time and robust requirements of the system.

W@WO3 Nanostructured films on carbon cloth was grown by the method of high temperature reactive thermal evaporation without catalyst and magnetron sputtering. The results of XRD, SEM and XPS show that the structure of the sample with good crystallinity, and the W particles adsorbed on WO3 nano film to form the structure W@WO3 Ohmic junction can be used to improve the charge separation efficiency of semiconductor materials. When the sample was exposed to Rhodamine B (RB) solution for 5 hours, 76% of Rb could be degraded. Compared with WO3 nanostructure with 26% degradation efficiency, the photocatalytic performance of the sample film modified by W could be greatly improved. In this experiment, carbon cloth was used as substrate to grow nano films with large specific surface area, which can provide more active sites for photocatalytic reaction. The experiment does not need expensive noble metal modification and catalyst. It can provide an effective preparation method for the preparation of high-quality photocatalytic materials with low cost, and has a broad application prospect in the field of wastewater treatment and degradation.

Lung lobes segmentation based on CT images is one of the most important references for doctors to diagnose and treat lung diseases. However, the blurred boundary of lung lobes and the huge workload of manual segmentation make it difficult for doctors to segment lung lobes accurately and quickly. To solve this problem, a new method of automatic segmentation of lung lobes based on 3D full convolution neural network is proposed. Firstly, the original CT image was preprocessed, then the convolution neural network was trained with the preprocessed image. Finally, the image to be segmented was input into the trained network model, and the lung lobe was segmented automatically. The experimental data included CT images of 50 patients with pulmonary diseases from Shanghai Pulmonary Hospital, 30 of which were used for training and 20 for testing. The segmentation results were quantitatively evaluated, with Dice coefficient of 0.961 and Jaccard similarity coefficient of 0.916. Experimental results show that the proposed automatic lung lobes segmentation algorithm has better segmentation performance and stronger generalization ability. It can segment the lung lobes accurately and quickly even when the training set data is small.

Diabetic macular edema is one of the main causes of blindness, the examination of OCT images by professional doctors is the main method to diagnose the DME, but this process is not only time-consuming but also prone to misjudgment. An auxiliary diagnostic model was proposed to discriminate the DME and normal macula. Firstly, the original OCT image is preprocessed by denoising, flattening and cropping to get the lesion area image which is easy to classify. Based on the pyramid model of wavelet decomposition, texture features are extracted from the original image and low-frequency sub-image by local binary mode method, and then fused with the gray-gradient co-occurrence matrix feature of the extracted detail image to form the final global feature, and reduce its dimension. Finally, the sequential minimal optimization algorithm of the weka platform was used to classify these images. The experimental results on Duke University and clinical datasets show that the accuracy, sensitivity and specificity of the proposed algorithm are 95.7% and 95.3%, 95.3% and 95.5% , 96.0% and 95.1% respectively. Therefore, the method can effectively classify OCT images and provide technical support for clinical auxiliary diagnosis of retinal diseases.

In view of the low accuracy of the current deep learning method in medical image registration, an unsupervised 3D convolutional neural network model for brain registration is proposed. The convolution network is used to regress the displacement field, and then the floating image is transformed through the spatial transformation layer. Then the network parameters are optimized according to the constructed loss function to realize the end-to-end unsupervised learning. By adding attention gate structure, low resolution auxiliary features are added to the connection between corresponding layers of the network to increase features and reduce background information. Compared with the unsupervised U-Net and VoxelMorph in MICCAI2012 multi-graph data, The results show that the method has higher registration accuracy and faster registration speed, and does not require expert annotation information, so it has good application potential in medical image registration.

Remote photoplethysmography (rPPG) can be used to calculate heart rate from face video. However, most of the proposed rPPG algorithms have many requirements and restrictions on subjects, which is not conducive to heart rate measurement under daily conditions. An improved method based on Plane-Orthogonal-to-Skin algorithm (POS) is proposed to solve the problem of unstable heart rate measurement due to the movement of the tested subjects in the actual environment. Using the frequency domain filter based on color distortion and POS algorithm, the signal is firstly filtered in the frequency domain, and then sent to the POS algorithm for time domain filtering, the influence of different noises on the signal quality is reduced. The result shows that the evaluation indexes of this algorithm are improved compared with other algorithms, and it is suitable for stable heart rate detection under face motion conditions.

In view of the problems existing in the traditional image description methods, such as the accuracy of extracting key information is not high and the description is not accurate, an image description method combining residual learning and dual-mode CAE is proposed. Firstly, a new dual-mode structure is proposed, which includes two inputs of image and text, as well as encoding, hiding layer interaction, decoding and other processing links to complete the text description of the input image. Then, residual learning is added to the classical convolution auto-encoder (CAE), and the convolution layer of CAE forms the residual neural network (DRN), which increases the learning depth and improves the accuracy of the method. Finally, the hidden layer of text and image is cross reconstructed to minimize the loss function, and the relationship between image and text is trained to realize the description of image. Using COCO and Flickr30k datasets to carry out qualitative and quantitative simulation experiments on the proposed method, the results demonstrate the effectiveness of the proposed method. Compared with other methods, the evaluation index Med r is the lowest, and R@K(K=1,5,10) was the highest, and the operation time is only 0.183s, which can describe the image more accurately than other methods.

Aiming at the problems of insufficient detail and over smooth in current face super-resolution algorithms, an algorithm for the Single Image Super-Resolution Reconstruction based on the Generative Adversarial Network(GAN) is proposed. The algorithm connects the edge detection network in parallel in the generation network, extracting abundant face contour details to assist in feature extraction, optimizing the network training process through the Ranger optimizer. Finally, establish a mathematical model to comprehensively evaluate the reconstruction effect combining objective assessment and subjective assessment indicators. The experimental results show that the algorithm has better subjective and objective effects than the Cubic Spline Method, SRGAN, FSRCNN, etc. It is proved that the algorithm improves the reconstruction ability of facial details and has a better reconstruction effect.

The infrared image of sulfur hexafluoride (SF6) has low contrast and blurred texture details and it is difficult to enhance leakage areas. An improved SF6 infrared image enhancement algorithm based on adaptive histogram equalization (CLAHE) with limited contrast is proposed. First, the original image is divided into basic image and detail image by bilateral filtering. Then the CLAHE algorithm is used to process the basic image and improve the contrast of the leakage area. And the detail image is transformed by piecewise linear transformation and Laplace transform image. It can edge image of highlighted image. Finally, the two images are linearly stacked to reconstruct the final infrared image, and the image enhancement is realized. The experimental results show that the enhancement effect of the algorithm on the leakage area of the SF6 infrared image is better than that of several common infrared image enhancement algorithms. It not only effectively suppress the noise and improve the contrast of the leakage area, but also highlight the edge of the leakage area and enrich the detail information.

In order to improve the accuracy of micro-expression recognition in video recognition field, a micro-expression recognition algorithm based on long short-term memory network and feature fusion is proposed. Both color features and texture features of micro expression images are extracted, and the selected spatial features are delivered to convolution neural networks for fusion. A new long short-term memory network is designed to learn temporal correlation between spatial features, the fused features are delivered to long short-term memory network to learn temporal features of micro expressions, the long short-term memory network is connected to classification network to recognize the class labels for each micro expression. Validation experiments are carried on two public micro expression recognition datasets, the results show that the proposed algorithm realizes better micro expression recognition performance, it realizes accuracy of 64.7% and 65.8% on SMIC dataset and CASME Ⅱ dataset, respectively.

In order to detect abnormal behaviors in surveillance video accurately and efficiently, a weak supervised abnormal behavior detection method based on improved yolov3 is proposed. Firstly, the multi-scale fusion method is used to improve the YOLOv3 network, and the improved yolov3 is used to complete the target detection in the video, which improves the computational efficiency and the universality of the method. Then, the large-scale optical flow histogram descriptor (LSOFH) is proposed to describe the target behavior by using the optical flow which can effectively capture the motion information, so as to better extract the abnormal behavior features. Finally, the least squares support vector machine (LSSVM) is trained to identify abnormal behaviors in surveillance video. Based on MATLAB simulation platform, the proposed method is verified by experiments. The results show that, compared with other methods, the proposed method performs best on the UCSD data set, UMN data set and subway exit data set, that is, the area under the curve (AUC) is the largest, the equal error rate (EER) is the lowest, and the detection rate is the highest. It has a good application prospect.