To improve the performance of Fourier transform spectrometer,a high-precision velocity information acquisition scheme based on equivalent clock method is designed.The laser signal forms an interference signal through the interferometer,which is amplified,filtered and shaped into a pulse signal recognized by the digital circuit.Based on the mathematical principle of velocity information acquisition,the error of obtaining velocity information based on T method is analyzed,and the velocity information acquisition method based on equivalent clock method is proposed.After reading the pulse signal,field programmable gate array (FPGA) calculates the optical path difference velocity value according to the equivalent clock method.Simulation analysis and experimental results show that when the frequency of He-Ne laser interference signal is 9 kHz,the measurement error of velocity feedback based on equivalent clock method is only 0.01%,and high precision optical path difference velocity information acquisition is realized.This is of great significance to improve the control accuracy of the interferometer system and the signal-to-noise ratio of the spectrum.

Traditional on-chip electrical interconnection has been unable to meet the increasing communication demands of multi-core processing systems.On-chip optical interconnection,which has advantages in terms of delay,energy consumption,and bandwidth,has gradually attracted attention.In order to reduce the hardware overhead of optical network-on-chip (ONoC) and improve the performance of the optical network,this paper proposes a 16-port passive H-tree optical interconnection network based on a microring resonator.The broadband micro-ring resonators is used to design 4 sets of steering optical routers,the use of micro-ring resonators is reduced and port selection is completed,and signals is transmitted to 8-port receiving optical routers and 3-level and 4-level optical switches to meet the contention-free transmission of signals.The experimental results show that compared with passive network structures such as Crossbar,λ〖ＷＴＢＺ〗-Router,GWOR,LACE,and Light,the passive optical H-tree network only needs 72 microring resonators at the 16×16 array scale.The average network insertion loss is 1.49 dB,which is reduced by 21.5%,10.7%,and 59.7% respectively compared with λ-Router,GWOR,and TAONoC.The average signal-to-noise ratio of each path reaches 17.48 dB,which is compared with λ-Router,GWOR,and Light respectively,increased by 38.5%,36.0%,and 17.1%.

Quartz glass is widely used in optical instruments of aerospace,medical exploration,optical imaging and its optical performance is one of the key factors that must be considered in the design of the instrument.The thermo-optical coefficient is an important optical performance parameter,and its change will affect the imaging quality and detection accuracy of the optical element.Therefore,the accurate measurement of the quartz glass thermo-optical coefficient is necessary for the design and calibration of the instrument.In this paper,a single-mode passive silica fiber is used as the measurement object,and a method for measuring the thermo-optical coefficient of quartz glass based on the all-fiber laser self-mixing interference technology is proposed.In the experiment,pure silica single-mode fiber is used as the external cavity of the interference system and fused with the fiber laser to form an all-fiber enclosed laser self-mixing interference thermo-optic coefficient measurement system.A fixed temperature range is given by the high-temperature heating box,and the thermo-optical coefficient of the quartz glass is calculated by the interference fringe counting method.The experimental results show that the resolution of the system to the thermo-optical coefficient can reach 1.632 9×10－7/(0.1 ℃).This research result lays a foundation for further research on the measurement of the thermo-optical coefficient of quartz glass and the application of fiber laser self-mixing interference.

Aiming at the high-speed data readout from the on-chip memory of the uncooled infrared detector,a low-latency sensitive amplifier for the on-chip memory of the uncooled infrared detector is designed.As the pixel array of uncooled infrared detectors continues to increase,the requirements for the on-chip memory of uncooled infrared detectors are also higher,and a higher-speed memory is needed for the internal data storage of the infrared detector.Reducing the delay time of the sensitive amplifier is a reliable method to increase the data transmission speed.In this paper,the traditional cross-coupling structure of the sensitive amplifier is improved.Compared with the traditional cross-coupling structure of the sensitive amplifier,a completely complementary second-stage cross amplifying circuit is added,and an intermediate stage composed of NMOS is used to couple the two-stage operational amplifier.The improved new sensitive amplifier can quickly and effectively amplify the voltage difference on the bit line,and at the same time improve the problem of low sensitivity.The sensitive amplifier designed in this paper uses TSMC 65 nm process.Under the conditions of 5 V working voltage and 100 mV bit line voltage difference,the simulation results show that the data readout delay is only 25.19 ps.Compared with the cross-coupled sensitive amplifier,the readout delay is reduced by 37.07%.At the sametime,under the simulation conditions of the full process angle,the ambient temperature is －45 ℃to 125 ℃,the delay simulation maximum value of the new sensitive amplifier is only 39 ps,and the minimum value is 17.1 ps.

In the application of biological tissue engineering,three-dimensional and longitudinal evaluation of biological structural markers is needed.However,these biological structures are usually several millimeters thick and cloudy,thus,it is very challenging for imaging,and classical fluorescence microscopy can not meet its needs.Therefore,we developed a mesoscopic fluorescence molecular tomography method.Mesoscopic fluorescence molecular imaging system was a new imaging system,which filled the gap between microfluorescence molecular imaging technology and macrofluorescence molecular imaging technology.In order to improve the performance of mesoscopic fluorescence molecular reconstruction,this paper mainly improved the configuration of the system.The configuration parameters of the optical system,including detector layout,uncoupled or coupled scanning mode,were optimized based on optical principles,and the 3d imaging performance of the mesoscopic fluorescent molecular imaging system was evaluated and compared.The results show that the coupling backlight mesoscopic fluorescence molecular tomography imaging (MFMT) system designed in this paper can improve the reconstruction performance and obtain high-quality reconstruction results.

Aiming at the problem of current slow detection speed of insulator defects in high-voltage lines and low accuracy in complex scenarios,this paper proposes an insulator defect detection algorithm for complex scenes based on lightweight you only look once (YOLOv4).Firstly,the lightweight efficient channel attention GhostNet (ECA-GhostNet) is used as the backbone to improve the detection speed.Then the classification-IoU joint representation is introduced in the head,and the general distribution is utilized to represent the flexible distribution of the bounding boxes to improve detection performance in complex scenes.In the training phase,quality focal loss (QFL) and distribution focal Loss (DFL) are used to better supervise joint representation and bounding boxes regression.Proposed method verifies the two types of targets of normal and self-explosive defective insulators on a dataset with complex background.The results shows that the detection accuracy of our approach in complex scenes is better than the current mainstream algorithms,and the detection speed reaches 49 FPS,which is about 40% higher than the original YOLOv4 algorithm′s detection speed.

Aiming at the problem of missed detection and poor detection effect of remote sensing images due to insufficient feature extraction and expression capabilities in complex backgrounds,a YOLOv4 algorithm model that optimizes feature extraction network is proposed.The improved model introduces a new Dense-PANet structure to obtain higher resolution features,and embeds the attention mechanism in the feature extraction network to adapt to remote sensing images due to the large field of view,which leads to the missed detection of small targets in complex backgrounds and the problem of poor detection results.In order to prove the effectiveness of the method proposed in this paper,a comparative experiment was conducted on DIOR remote sensing data sources.The results show that the average accuracy (mean average precision,mAP) of the algorithm in this paper is 86.55%,which is an increase of 2.52% compared to the original algorithm.YOLOv3 and RetinaNet increased by 6.58% and 14.09%,which verifying the effectiveness of the improved algorithm.

The detection of remote sensing images has a wide range of applications in monitoring the natural environment,military,homeland security and so on,while remote sensing images have the disadvantages of complex background,small target area and difficulty in character extraction.In this paper,a remote sensing image detection algorithm based on selective fusion of multi-scale features is proposed.The proposed algorithm uses the improved Resnet50 as the backbone network,replaces the first convolution of the Resnet50 with dynamic convolution,and replaces the convolution in the ConvBlock module with pyramid convolution to improve feature extraction capability.At the same time,in order to avoid missing the underlying information,the proposed effective spatial channel attention mechanism module is added after the dynamic convolution layer.Finally,the different scale features based on context information are selected to fuse and improve the model′s ability to locate the target object.The experimental results show that the algorithm improves the detection accuracy of remote sensing images while ensuring speed,and the mean average precision (mAP) reaches 91.88% and 90.23%,respectively,on the remote sensing image disclosure data set RSOD and NWPUVHR-10,and thedetection speed reaches 33 FPS.

In recent years,convolutional neural networks have been widely used in the field of image super-resolution.The super-resolution algorithm based on convolutional neural network has some problems,such as insufficient feature extraction of image,large number of parameters and difficult training.Therefore,this paper proposes a lightweight image super-resolution reconstruction algorithm based on gated convolutional neural network (GCNN).Firstly,the shallow feature extraction of the original low-resolution image is carried out by convolution operation.Then,the gated residual blocks (GRB) and long and short residual connections fully extract image features,and its high-efficient structure can also accelerate the network training process.The gated unit (GU) in the GRB uses the regional self-attention mechanism to extract the weight of each feature point in the input feature map.And then it multiplies the gate weight by the input feature element by element as the output of the GU.Finally,high-resolution images are reconstructed using sub-pixel convolution and convolution module.Experiments are conducted on Set14,BSD100,Urban100 and Manga109 datasets.Compared with the classical methods,not only does the algorithm in this paper have higher peak signal-to-noise ratio and structural similarity,but also the reconstructed image has clearer contour edges and details.

In order to realize the fast and accurate appearance quality classification of picked fruits,and cooperate with the sorting production line to complete the large-scale centralized sorting of fruits,a fruit classification method based on improved ResNet is proposed in this study.Firstly,the residual module in ResNet network is combined with the dual channel squeeze-and-excitation block (DC-SE Block) to enhance the effective channel features,suppress the inefficient or invalid channel features,and improve the expression ability of the feature map,so as to improve the recognition accuracy.Secondly,the Inception module is added to the original ResNet model to fuse the characteristics of different scales of fruit,so as to enhance the recognition ability of small defects.Finally,four kinds of fruit images with different appearance quality are enhanced,and the model is initialized by transfer learning method.Taking apple as an example,the experimental results show that the accuracy of the improved model trained by the data set is 99.7%,which is higher than 98.5% of the original model;The precision rate is 99.7%,which is higher than 98.3% of the original model;The recall rate reaches 99.7%,which is higher than 98.7% of the original model; The average detection speed under graphic processing unit (GPU) is 32.3 frame/s,which is slightly lower than 35.7 frame/s of the original model.Compared with several advanced classification methods such as GoogleNet and MobileNet,and compared with different improved models,the results show that the proposed method has good classification performance,and has important reference value for solving the problem of accurate classification of fruit appearance quality.

Facial expression recognition plays an important role in artificial intelligence such as human-computer interaction.However,current researchers ignore the semantic information of human faces.In this paper,we propose a facial expression recognition network fusing local semantic and global information,which consists of two branches:the local semantic region extraction branch and the local-global feature fusion branch.Firstly,the face semantic parsing is achieved by training semantic segmentation network on face parsing dataset.The semantic parsing of facial expression dataset is obtained by transfer training.Then the meaningful regions and their semantic features are extracted and fused with the global features to obtain the semantic local features.Finally,the global semantic composite features of facial expressions are constructed by combining semantic local features with global features.They are classified into one of the 7 basic facial expressions by the classifier.We also propose a training strategy of unfreezing partial layers,which makes semantic features more suitable for facial expression recognition and reduces the redundancy of semantic information.The average recognition accuracy on two public datasets,JAFFE and KDEF,reaches 93.81% and 88.78%,respectively.The performance outperforms the current deep learning methods and traditional methods.The experimental results demonstrate that the network proposed can describe the expression information comprehensively by integrating local semantic and global information.

Aiming at the problem of low image recognition rate in finger-vein recognition due to insufficient training samples,a finger-vein recognition method combining linear regression classification (LRC) and multi-sample expansion is proposed.First,the matrix transformation is used to generate a mirror image of the original image,all the original images and mirror images are trained, and the useful information contained in the finger-vein image is increased.Then,the test and training samples are classified based on LRC.Finally,the final classification result is obtained by calculating the deviation,and the recognition rate is found out.In addition,a finger-vein acquisition device is designed to collect and obtain a self-built finger-vein database.The experimental results show that the recognition rate of the proposed algorithm on the finger-vein database of the self-built database,the finger-vein database of Shandong University and Malaysian University of Technology reached 98.93%,98.89% and 99.67%,and the lowest error rate was 2.388 8%.Compared with other traditional and popular algorithms,the experimental results have obvious advantages and good practical application value.

Medical X-rays,as a routine examination method for chest diseases,can diagnose early and unobvious chest diseases and observe the lesions.However,the characteristics of multiple diseases on the same radiographic image are a challenge to the classification problem.In addition,there are different correspondences between disease labels,which further leads to the difficulty of classification tasks.In response to the above problems,this paper combines the graph convolutional neural network (GCN) with the traditional convolutional neural network (CNN),and proposes a multi-label chest radiographic image disease classification method that combines label features with image features.This method uses the graph convolutional neural network to model the global correlation of the labels,that is,constructs a directed relationship graph on the disease label,each node in the directed graph represents a label category，and then inputs the graph into the graph convolutional neural network to extract the label features,and finally merges with the image features to sort.The experimental results of the method proposed in this paper on the ChestX-ray14 dataset show that the average AUC of 14 chest diseases reaches 0.843.Compared with the current three classic methods and advanced methods,the method in this paper can effectively improve the classification performance.