A low polarization-dependent Ge/SiGe multiple quantum wells electro-absorption modulator by introducing biaxial tensile strain through suspended microbridge structure is fabricated on complementary-metal-oxide-semiconductor (CMOS) compatible silicon platform using ultraviolet lithography technique (UVL).The biaxial tensile strain of the fabricated device is measured through Raman spectroscopy,and the photocurrent response,modulation extinction ratio and high frequency response of the device are tested under both of transverse electric (TE) and transverse magnetic (TM) polarization.The device presents a low polarization-dependent extinction ratio of 5.8 dB under 0 V/4 V operation,together with a 3 dB modulation bandwidth of 8.3 GHz at 4 V reverse bias voltages.Compared with electron beam lithography technique (EBL),although the device fabricated through UVL performs a little worse in modulation extinction ratio and high frequency response bandwidth,it has the advantages of short exposure time,low cost and enabling mass production,which has broad application prospects.

Owing to its high photosensitivity and high accurate alarm rate,gallium oxide based solar blind photodetectors attract much attention in military and civilian fields.To meet the requirements of the modern device,gallium oxide based photodetector desired to be flexible and self-powered.However,there are few reports about Ga２O３ self-powered flexible photodetectors.In this work,we grew β-Ga２O３ nanowire on a flexible glass fiber fabric substrate and constructed an Ag/β-Ga２O３ Schottky junction.Thanks to the build in field,the Ag/β〖ＷＴＢＺ〗-Ga２O３ photodetector can work without an extra power.Meanwhile,the photodetector exhibited an ultra-low dark current (0.2 pA at 0 V),a high light-to-dark ratio (~500),a fast response speed (0.46/0.41 s) and a high detectivity of 6.8×109 Jones.The device also exhibited good mechanical properties.The results show that Ag/〖ＷＴＢＸ〗β-Ga２O３ Schottky junctions have promising applications in the field of deep-UV detectors and provide a new idea for the development of flexible self-powered detection technology.

A Fabry-Perot (F-P) interferometric fiber sensor based on femtosecond laser direct writing is designed to achieve dual parameters measurement of displacement and external strain,and the implementation principle of the sensor is studied and experimentally verified.First,a femtosecond pulsed laser is used to inscribe two lines of length 48 μm in the longitudinal direction of the core of a standard single mode fiber (SMF) with a spacing of 105 μm to form an F-P resonant cavity,and then one end of this structure is cut flat to form a composite interferometer together with an all-reflective mirror,thus constituting a displacement sensor.The cut flat end is then knotted to allow for stress measurements again.The experimental results show that the displacement sensitivity of the structure at the Dip1 position is －440.3 pm/μm with a linear fitting coefficient of 0.990 3;the strain sensitivity at the Dip2 position can be measured after the structure is knotted at 1.2 pm/με with a linear fitting coefficient of 0.992 4.The sensor is simple to fabricate,low cost,tiny structure,good linearity and easy to repeat.The sensor is of practical value in both displacement sensing and strain sensing fields.

For hyperspectral remotely sensed images,a subpixel mapping method of constrained spatial spectrum is proposed.Traditional subpixel mapping methods may not make full use of the rich spectral information of hyperspectral images due to straightforward using of unmixing results.The method based on constraint spatial-spectral subpixel mapping (CSSSM) is proposed,which explicitly combines the pixel abundance with the subpixel abundance through using subsampling,and derives the new subpixel abundance model by substituting the linear unmixing model.After adding sparsity and smoothness constraints to control the searching space of solutions,the subpixel abundance is obtained quickly.The reweighted 1-norm constraint is applied to the subpixel sparse abundance,and the weight is updated adaptively.Total variational （TV） regularization is used as a new constraint for a spatial priori of the subpixel abundance,and a multiplication iterative algorithm is used to search for the subpixel abundance.Finally,a winner-take-all strategy is used to determine the overall class categories.Experiments on two synthetic data sets show that the proposed method can improve the accuracy of subpixel mapping further.

Combined with the local features of point cloud and Octree optimization search,an automatic registration algorithm of 3D deformation point cloud for machining process measurement of thin-walled parts is proposed,and the displacement deviation is effectively calculated.Firstly,the data of the point cloud model of thin-walled parts is preprocessed to remove the invalid points and noise points in the main body.The normal vector and three feature elements of the point cloud are calculated as the input of the point pair feature net (PPFNET) feature learning method.The deformed local features are aggregated into the global features by using the maximum pool layer.Through the in-depth learning of the global and local feature descriptors,it can find out the corresponding relationship between disordered point clouds and complete the rough registration process of point clouds.Then,an improved precision registration algorithm based on interative closest point (ICP) is proposed.By increasing the threshold limit and filtering the influence of chatter during machining deformation,the registration accuracy is 98.58% and the registration efficiency is improved by 10%.Finally,Hausdorff is used to calculate the distance,and Cloud-Compare is used to analyze the displacement deviation.The comparison between the analysis results and the experimental results shows that the mean absolute percentage error (MAPE) is 2.32%.The simulation results show that the proposed method meets the requirements of real-time and measurement accuracy of machining deformation.

A multi-scale hybrid attention capsule network (CapsNet) model is proposed to solve the problem of insufficient feature extraction due to single feature extraction structure in CapsNet and excessive amount of parameters in data processing.First,convolution kernels of different scales are added at the initial end of the network to extract features at multiple angles,and channel attention (CA) mechanism and spatial attention (SA) mechanism are introduced to reduce complex background interference by focusing on features of more resolved regions.Second,a local pruning algorithm is adopted to optimize the dynamic routing algorithm,which reduces calculation parameters and training time.Finally,validation on open marine fish data set F4K (Fish 4 Knowledge) shows that the model recognition accuracy in this paper is 98.65% compared with traditional residual network50 (ResNet-50),bilinear convolutional neural network (B-CNN),spatial transformation network and hierarchical compact bilinear pooling (STN-H-CBP) and CapsNet models,5.92% higher than ResNet-50 model;The training time is 2.2 h,which is nearly 40 min shorter than that of CapsNet,which verifies the feasibility of the proposed algorithm.

Automatic detection of pavement defects is of great importance for road maintenance and road condition rating assessment.To this end,a pavement defect detection system was designed using YOLOv5x combined with perspective transformation and image segmentation.First,a multi-type pavement defect dataset (PDD) was collected and produced to demonstrate the feasibility of the system.Then,four models single shot multibox detector (SSD),Faster R-CNN,you only look once v5x (YOLOv5x) and YOLOX,were used to train the PDDs for detection.After training,the mean average precision (mAP) of all four models exceeded 77%,with YOLOv5x showing the best results with 91% mAP,while proving the validity of the created dataset PDDs.Finally,YOLOv5x was used as the main detection method of the system combined with perspective transformation,image segmentation and skeleton extraction to obtain information such as length,width and area of defects,and then calculating the pavement condition index (PCI) to obtain the pavement damage level and the corresponding repair suggestions,improving the practicality of pavement defect detection.

Underwater fish images are affected by ligt scattering and absorption,water impurities and other factors,resulting in low underwater fish image quality.This article uses improved automatic color equalization (ACE) algorithm to enhance underwater fish images to effectively improve image quality,and lay a good foundation for the subsequent underwater image segmentation.Aiming at the problems of poor segmentation effect and low real-time performance of underwater fish images,this paper proposes the ARD-PSPNet network model,using the ResNet101 network model as the feature extraction network,and using the pyramid scene parsing network (PSPNet) network model with good segmentation performance as the basic image The segmentation model reduces the amount of calculation by introducing deep separable convolutions.Through the R-MCN network structure,it makes full use of the rich location information and completeness of the shallow network feature layer,and improves the loss function to make the segmentation position more accurate.In experiments and completed on the Fish4knowledge data set.Experimental results show that the new model has an increase of 2.8% in mean intersection over union (MIOU) and about 2% in mean pixel accuracy (MPA) compared with the original model.

A scheme of single sideband vector millimeter wave signal generation and transmission based on double parallel Mach-Zehnder modulators (DP-MZM) is proposed.Under the joint simulation environment of MATLAB and VPI,the generation and transmission performance of different modulation formats in 90 GHz band are analyzed.In the simulation,when the DP-MZM works in the optical carrier suppression (OCS) mode,the single sideband signals of the two submodulators driven by small signal superimpose each other without the use of filters,the center carrier can be completely suppressed.The results show that the bit error rate (BER) of each modulation signal can reach the hard decision forward error correction (HD-FEC) threshold.The error performance of the quadrature amplitude modulation signal after probability shaping is obviously better than that of the uniform quadrature amplitude modulation signal,and the power gain of the 7 bit/symbol probility shaping 256 quadrature amplitude modulation (PS-256QAM) signal after systematic transmission is the most obvious,the power gain of 2.94 dB is increased after transmission of 50 km optical fibers.Compared with the existing schemes of carrier supresion millimeter wave,the system has the advantages of simple structure,high adjustability and obvious signal transmission gain,so it has certain application value.

In order to solve the problem of peak to average power ratio (PAPR) in DC-biased optical orthogonal frequency division multiplexing (DCO-OFDM) visible light communication (VLC) system,a joint PAPR reduction algorithm based on combining clipping and normalized μ-law companding is proposed.Firstly,some big peak signal in the time-domain is clipped by amplitude limiting to reduce the amplitude of the signal,and then the clipped signal is further processed by a normalized μ-law companding transform.With this way,the PAPR of DCO-OFDM signal can be effectively reduced.Experimental simulation result shows that the value of complementary accumulative function is １０－３ when 16QAM modulation is adopted,compared with the algorithm combined amplitude limiting with traditional μ-law companding transform,the PAPR of the combined algorithm decreases by about 2.206 dB.

Keratoconus causes the central cornea to bulge forward during the disease process,giving the cornea a conical shape,and leading to highly irregular myopia and astigmatism,causing damage of vision with different degrees.The disease generally occurs in the adolescent period,in order to timely treat and to avoid serious lesions, it is of great significance to screen and distinguish keratoconus.In addition,clinical diagnosis of keratoconus is usually detected by corneal topography,which can obtain morphological changes of the cornea,but there is a certain misdiagnosis rate.At present,it has been found that the change of mechanical properties of cornea is prior to morphology.Therefore,from the perspective of corneal biomechanics,this paper proposed a model to distinguish keratoconus based on multi-layer perceptron (MLP) neural network.Firstly,corneal visualization scheimpflug technology (Corvis-ST) was used to measure the biomechanical video of cornea,and corneal biomechanical parameters were obtained as a data set,including normal cornea and keratoconus.Then,MLP neural network model was constructed for corneal biomechanical parameter data sets,in which 70% data sets were used as training sets and 30% as test sets.The results of training and testing on the datasets showed that the accuracy of keratoconus differentiation was 97.6%.

In view of the problem that quality of pseudo-labels is uneven in the current retinal vessel segmentation task and it requires to be screened to obtain the high-quality pseudo-labels,a novel semi-supervised deep learning framework for retinal vessel segmentation is proposed in this paper.The framework adopts the idea of divide and conquer to process data.Traditional deep learning methods are utilized especially for the labeled data,while Mean teacher model is used to deal with the unlabeled data.By comparing the different morphological outputs of the same input,the model can learn the common features between the unlabeled data and avoid the screening process brought by pseudo-label technology.Three benchmark networks,u-neural networks (U-Net),Dense-Net and Ladder-Net are put into the framework,the experiments are carried out on DRIVE and CHASEDB1 datasets,which achieve good segmentation results.It shows that the framework can improve the ability of the network to distinguish different threshold pixels.

Aiming at the problems of low segmentation accuracy caused by the loss of network model information,insufficient context information and poor network generalization ability in brain tumor image segmentation,a new brain tumor image segmentation method is proposed.This method is a multi-level connected (MC) brain tumor segmentation model composed of depth gate convolution module (DGC) and feature enhancement module (FEM).The depth convolution module is used to reduce the information loss of feature information transmitted layer by layer.The control gate unit (CGU) is used to realize the MC of each scale feature map,in which the combination pooling is used to reduce the information loss in the down sampling process.The feature weight of the segmented region is enhanced by the FEM.The experimental results show that the Dice index of the whole tumor area (WT),tumor core area (TC) and enhanced tumor area (ET) predicted and segmented brain tumors reaches 0.92,0.84 and 0.83 respectively,and the Hausdorff distance reaches 0.77,1.50 and 0.92.Compared with many current methods,the segmentation accuracy and calculation efficiency of brain tumors are higher,and have good segmentation performance.

In order to address the limitations of single modal brain images and further highlight detail features and enhance visual effects,an algorithm framework based on multi-scale edge preserving decomposition and improved sparse representation (ISR) is proposed.First,the source image is decomposed to obtain high frequency and low frequency subbands.Secondly,an improved sparse representation with multi-norm weighted metric is used to fuse low-frequency subbands,and an improved guide filter with multi-scale morphological gradient (MSMG) is used to remove details.At the same time,the simplified pulse-coupled neural network fuses its high frequency subbands.Finally,the inverse transformation yields the fused brain image.Experimental results show that this paper has significant advantages in the protection of edge information,improvement of fusion efficiency and saving of time cost.