Metalens is a device based on metasurfaces and metagrating,which can flexibly regulate the phase,polarization,and amplitude of light. It has the advantages of lightweight and easy-integration.However,the fabrication cycle time of metalens is long and the cost is great,and it is extremely necessary to find an easy-to-process,low-cost and high-efficient method.In this work,an all-dielectric terahertz (THz) metagrating with highly efficiency wavefront control in THz range is proposed.The all-dielectric metagrating bends the electromagnetic waves beam into the T-1 diffraction order when the electromagnetic wave is incident normally. In the simulation,more than 83.44% of the transmitted energy is concentrated in the T-1 diffraction order for P-polarized light and more than 82.73% for S-polarized light.Then,based on the all-dielectric metagrating,a metalens with numerical aperture of NA=0.39 at 0.14 THz is proposed,the metalens has been fabricated by 3D printing.The focusing capability of the metalens is characterized by utilizing the scanning measurement system.The focal spot with the focusing distance of 114.5 mm is achieved,the measured and anticipated results are satisfactorily congruous with one another,which could validate our design.Meanwhile,the size of the focal spot is obtained.The transmission THz imaging system is performed to characterize the imaging capability of the metalens.This work holds great promise for extensive applications in light sensing,communication,and super-resolution imaging.

A microwave photon frequency conversion method is put forward and demonstrated using Mach-Zehnder modulator (MZM) and injection-locked laser.In the upper branch,the laser is modulated by the radio frequency (RF) signal to realize the carrier suppression modulation and then the optical sideband is filtered by the filter;in the lower branch,the laser is modulated by the local oscillator (LO) signal to realize single sideband modulation and then filtered and amplified by the injection-locked laser.After the coupling of the two beams,the light enters the photodetector (PD) beat frequency to realize up and down frequency conversion.The structure makes full use of the feature that the injection-locked laser can be locked by the input laser,and can obtain perfect local oscillator light, so it can generate better microwave signal through the difference frequencies.Simulation results show that the spurious free dynamic range (SFDR) of the system is greater than 90 dB·Hz2/3 and the noise figure (NF) is less than 30 dB.

The Bloch surface wave (BSW) has the properties of low transmission loss,significant local field enhancement at the surface as well as sensitive to the environment,which has been regarded as one of the key solutions for the study of near field light-matter interaction.BSW mainly propagates along the truncated one-dimensional photonic crystal and the external medium,the manipulation of BSW can be realized at the nanometer scale via the micro-nano structural design of the propagation interface.Here,embarking from the optical properties of BSW mode,a graphene loaded one-dimensional photonic crystal BSW sensor structure is proposed.The existence of the excited mode in the band-gap can be determined by tuning the structural parameters of the graphene layer.Through investigating the optical field transmission characteristics of the excited mode,the manipulation of the excited wavelength,intensity as well as phase of the sustained BSW mode can be achieved.Moreover,leveraging its sensitivity to the refractive index change of the external medium,the sensing detection capability of this proposed structure is discussed.The results demonstrate that the proposed sensor structure is expected to achieve highly sensitive biochemical sensing detection applications.The study provides a new route for the design and development of new BSW integrated photonic devices.

The hot-rolled steel strip surface defects detection methods have low detection accuracy and suffer background clutters.To solve the issues mentioned above,this paper proposes a CA-YOLOv5 defects detection method based on coordinate attention (CA).The improvement of CA-YOLOv5 are three aspects.First,to make the YOLOv5 more suitable for strip defects detection,the data argumentation using the 4 or 9 mosaic images randomly and hyperparameter optimization based on the genetic algorithm (GA) are employed in the input of the network.Second,the CA mechanism is introduced between the backbone and the neck model to improve the feature extracting ability of the network.Third,each detecting head is decoupled to separate the classification and regression tasks,which could further boost the accuracy of defects detection.Comparative experiments on NEU-DET dataset show that the proposed network obtains 84.36% mean average precision (mAP),which not only exceeds YOLOv5 by 6.68%,but also outperformances other state-of-the-art detectors.

Due to the complex background of flower images taken under natural conditions and their high intra-class variability and inter-class similarity,it is difficult to achieve accurate fine-grained classification by existing popular methods relying only on the convolution module to extract local features of flowers.To address the above problems,this paper proposes a high-precision and lightweight flower classification method (ConvTrans-ResMLP).It achieves global feature extraction of flower images by combining the Transformer module and the residual multi-layer perceptron (MLP) module,and adds convolutional computation to the Transformer module so that the model still retains the ability to extract local features.Meanwhile,in order to further deploy the model to edge devices,this study achieves compression and optimization of the model based on knowledge distillation.The experimental results show that the accuracy of proposed method achieves 98.62%,97.61% and 98.40% on Oxford 17,Oxford 102 and homemade Flowers 32 datasets,respectively.The size of the lightweight model in this paper is about 1/18 of the original one after knowledge distillation, while the accuracy rate only decreases by about 2%.Therefore,this study can better improve the efficiency of flower fine-grained classification by edge equipment, which is of practical significance to promote the automation of flower cultivation.

Manual surface detection of cable-stayed bridge cables is low accuracy and high labor-intensive.The speed of conventional image processing and convolutional neural networks is too low to meet the requirements for timely detection.Therefore,a novel encoder-decoder network is constructed to detect cable surface defects.The optimized MobileNetV2 is used as the encoder to reduce the model parameters and increase the training speed.The UNet idea and pyramid pooling (PSP) module are used in the decoder to enhance the feature extraction.Moreover,skip connections connect the encoder and decoder to fuse the deep and shallow feature information effectively.The PASCAL VOC dataset is used to pre-train the network to obtain the weight values of the network, which are then loaded into the network to obtain the final parameters through the training of defect datasets such as holes,gaps and damages.The experiments demonstrate that the novel encoder-decoder network is robust.The mean pixel accuracy,mean intersection over union and the processing time of single image are 89.88%,79.25% and 41.34 ms respectively,which are better than the methods,such as PSPNet,UNet and DFANet. In summary,the novel network meets the requirements of accuracy and speed for surface defect detection of cable-stayed bridge.

Road extraction from remote sensing images is of great significance in promoting urban and rural development planning and construction.However,the traditional methods for road extraction from remote sensing images have the problems of large engineering quantities and low efficiency,and the methods based on depth learning have the problems of low extraction accuracy and poor connectivity in complex scenes.To solve the above problems and improve the accuracy of road extraction in different geomorphic regions,this paper proposes a road extraction method based on iHDODC LinkNet network for high-resolution remote sensing images.This method is improved on the basis of the semantic segmentation model D-LinkNet:on the one hand,ResNeSt50 is used to reconstruct the D-LinkNet network and a pre training model is added to propose a hybrid depthwise over-parameterized dilated convolution (HDODC) module.On the other hand,iterative attentional feature fusion (iAFF) mechanism is used to replace the original additive fusion,so that the model focuses on the global information of the road.Finally,the training is carried out on the Massachusetts road dataset and a provincial highway scene dataset,and the effectiveness of the improved model is proved by the extraction effect of the test set.According to the experimental model segmentation effect,the improved method applied to F1 reaches 71.66%,which is 10% higher than the original model,and better segmentation results can be obtained.

To realize separation and phase demodulation of superimposed interference intensity signals in multi-surface interferometry measurements,a frequency-corrected-based multi-surface wavelength-shifting interferometry algorithm is proposed,which can realize the simultaneous reconstruction of the surface shapes of the measured transparent plates.The wavelength-shifting interferometry technique can enable each interference harmonic to feature different phase-shifting values according to the corresponding optical path difference (OPD),which provides the basis for the separation and phase demodulation of each signal.In existing multi-surface measurement techniques,the harmonic frequencies are often roughly estimated from information such as the cavity length and optical thickness of the measured plates, but the estimation accuracy is low and cannot cope with phase-shifting errors.Therefore,this paper realizes the accurate extraction of each interference harmonic frequency by multi-point averaging and frequency correction,which can effectively eliminate the influence of abnormal value and additive Gaussian noise (AGN) on the frequency solution accuracy,and can demodulate each harmonic phase simultaneously only by the weighting operation between interferograms.The comparative analysis and experimental results verify the reliability of the proposed algorithm.

To address the issues that the detection rapidity and accuracy are disturbed by the tiny defects,different scales and other factors on the surface white print of the mobile phone chip shielding shell,an long short link and double attention network (LSDANet)-based surface defect detection method is devised in this paper.First,the feature extraction ability of the network model for defects with different scales is enhanced via constructing an encoding and decoding-based semantic segmentation model and utilizing the long short-distance connection path.Second,the feature weights of white print defects with a size of 5 to 10 pixel in space and channel are increased via designing the space-and channel-based attention mechanisms,respectively.Ultimately,a LSDANet defect detection network using the dual attention mechanism and long short-distance connection path segmentation model is proposed for surface defect detection of the mobile phone chip shielding shell.The experimental results demonstrate that the detection performances of the LSDANet-based algorithm in mean pixel accuracy,mean intersection over union and frames per second are 96.21%,66.13% and 39.03,which are superior to the other semantic segmentation methods in terms of detection precision and speed.

Aiming at the low detection precision of small objects and noise interference in light detection and ranging (LiDAR) point cloud 3D object detection based on deep learning,a 3D point cloud object detection method CSA-RCNN（cross self-attention region cnn）based on cross self-attention mechanism was proposed.The cross self-attention was used to learn the coordinates and features of the point cloud simultaneously,and a multi-scale fusion (MF) module was designed to adaptively capture multi-scale features at each level.In addition,an overlapping sampling strategy was designed to selectively resample the target region of interest to obtain more foreground points,effectively reducing noise sampling.The algorithm performance test was carried out on the widely used KITTI dataset.The results show that the detection precision of the method in this paper for small objects such as pedestrians is greatly improved,and the average precision mean value is increased compared with four classical algorithms such as PointRCNN,which significantly improves the performance of 3D point cloud object detection.

Based on the generalized Huygens-Fresnel principle and the theoretical model of atmospheric turbulence,the analytical expression of the spiral spectrum and the expression of the light intensity of the Gaussian vortex beam in the non-Kolmogorov atmospheric turbulence with a slant path are derived,and the distribution rules of spiral spectrum and light intensity after the transmission of the vortex beam are numerically simulated.The effects of each beam parameter and atmospheric turbulence on the spiral spectrum dispersion are analyzed.The results show that with the increase of the transmission distance,the dispersion of the spiral spectrum is stronger.When the transmission distance increases to a certain distance,the spiral spectral components gradually tend to be uniform distribution, and the light intensity gradually presents a Gaussian distribution.Increasing the initial topological charge and wavelength of the beam can effectively reduce the dispersion degree of the spiral spectrum after transmission.When the zenith angle gradually increases to π/2,the transmission mode is horizontal transmission,and the dispersion degree of spiral spectrum is significantly increased.The larger the refractive index structure constant at the ground and the smaller the inner scale,the more serious the dispersion of the spiral spectrum after transmission,while the influence of the outer scale on the spiral spectrum is very small.

In order to visualize the process of laser polishing metal surface,a two-dimensional transient numerical model for pulsed laser polishing of 304 stainless steel is established in this paper.The evolution process of surface morphology of 304 stainless steel under the process of pulsed laser is simulated,and the evolution mechanism of heat transfer,melting,solidification and other physical processes in the polishing process is revealed.The simulation results show that compared with continuous laser polishing, pulsed laser will produce fine ripples on the polished surface,but the overall spatial curvature will be reduced,which can play a certain polishing effect. In the process of surface morphology evolution,the capillary force dominates the melt flow in the molten pool in the area with large surface undulation,and its main role is to eliminate the surface morphology with large area ratio;On the contrary,the thermal capillary force dominates,and its main role is to promote the redistribution of molten metal in the molten pool and further reduce the surface roughness.In order to verify the correctness of the simulation model,by comparing with the experimental results,it is found that the remelting zone depth of the model is consistent with the experimental results,and has a high simulation accuracy.

Aiming at the problems of low utilization of feature information and insufficient generalization ability in the traditional medical image segmentation network with encoding and decoding structure,this paper proposes a multi-scale semantic perceptual attention network (MSPA-Net) combined with encoding and decoding mode.Firstly,the network adds a dual-channel multi-information domain attention module (DMDA) to the decoding path to improve the ability of feature information extraction.Secondly,the network adds a dense atrous convolution module (DAC) at the cascade to expand the convolution receptive field.Finally,based on the idea of feature fusion,an adjustable multi-scale features fusion module (AMFF) and a dual self-learning recycle connection module (DCM) are designed to improve the generalization and robustness of the network.To verify the effectiveness of the network,the experimental verification is carried out on CVC-ClinicDB,ETIS-LaribPolypDB,COVID-19 CHEST X-RAY,Kaggle~~3m,ISIC2017,and Fluorescent Neuronal Cells datasets,and the similarity coefficients reach 94.96%, 92.40%,99.02%,90.55%,92.32% and 75.32% respectively.Therefore,the new segmentation network shows better generalization ability,the overall performance is better than the existing network,and can better achieve the effective segmentation of general medical images.