On the basis of Mie scattering theory and Taylor′s frozen hypothesis,a dual laser non-Doppler lidar wind measurement system is developed in order to design a low-cost,high-precision,telemetered wind speed measurement device.A 532 nm laser is used as the light-emitting unit,a charge-coupled device (CCD) as the light-receiving unit,and a computer as the information processing unit in the design of the experimental non-Doppler lidar wind measurement device.Under conditions of 3 m/s and 2 m/s wind speed,the foward scattered echo signals of the dual laser beams produce the light intensity maps of the left and right lasers,respectively.After analyzing the light intensity map,it is possible to calculate how long it would take for an aerosol to migrate from one laser beam to another.Using the distance between the two laser beams,the wind speed measurements are obtained under the conditions of wind speeds of 3 m/s and 2 m/s.The errors are 7% and 7.33%,respectively,between the mean measured value and the actual value.The non-Doppler lidar wind measurement device is simple in design,affordable,and has a significant practical application.

Aiming at the detection accuracy and speed of pedestrian detection in complex road environment,a lightweight pedestrian detection algorithm based on multi-scale information and cross-dimensional feature guidance is proposed.Firstly,based on the high-performance detector YOLOX,a multi-scale lightweight convolution is constructed and embedded in the backbone network to obtain multi-scale feature information. Secondly, an end-to-end lightweight feature guided attention module is designed,which guides the model to focus on the visible region of pedestrian targets by fusing spatial information and related information through cross-dimensional channel weighting mehod. Finally,in order to reduce the loss of feature information in the process of lightweight of the model,a feature fusion network is constructed by depthwise separable convolution with increasing the depth of the receptive field.The experimental results show that compared with other mainstream detection algorithms,the proposed algorithm on the KITTI dataset reaches 71.03% detection accuracy and 80 FPS detection speed,which has better robustness and real-time performance in scenes with complex background,dense occlusion and different scales.

Most of the existing bridge crack image inpainting methods are single target restoration,which cannot generate multiple reasonable filling contents based on valid information around the hole. Moreover,the inpainting results suffer from structural distortion and texture blurring.A diversity crack image inpainting network based on the mask distance convolutional block attention module (MD-CBAM) is proposed in this paper,which mainly consists of a diversity structure generator and a texture generator.The regional structure attention is proposed to reduce the difference between the pixels in the masked region and the valid pixels,and the average pooling is performed on the attention scores according to the mask features to improve the inference ability of the model to the masked area.The MD-CBAM module is designed to synthesize high-quality features in the texture generation stage.The module utilizes the distance information between features and semantic information to effectively enhance the capability of the model to fill large holes.The experimental results show that the inpainted image has a more definite structure and a more reasonable texture,and the peak signal-to-noise ratio (PSNR) and Fréchet inception distance (FID) reach the best at each mask ratio, where the PSNR increases by 0.22—2.38 dB at the mask ratio of ［0.4,0.5) and the structural similarity (SSIM) value is optimal.

Aiming at the problems that the existing dehazing algorithms do not fully consider the fog information of haze image and the blurred details of the restored image,a novel haze feature map reflecting the distribution of fog information is proposed, and an inequality relation constraint method is adopted to enhance image quality.First,the extreme value channels of degraded image are extracted to achieve a rough estimation of fog information, and optimized by using L-1 regularization to obtain a haze feature map.After that,a primary atmospheric light veil function based on haze feature map is presented.Through in-depth analysis of color channel and atmospheric light veil, the constrained atmospheric light veil is obtained by using the mean value inequality.Finally,the local atmospheric light is improved by using the haze feature map,and haze removal is achieved based on the atmospheric scattering model.Compared with other existing state-of-the-art methods on both real-world and synthetic datasets haze images,our method shows favorable performance for single image dehazing especially in night image dehazing.

The compressive light field (CLF) display using multi-layer spatial light modulators has the advantages of high space bandwidth utilization rate and good image resolution.It is a promising light field display technology.The traditional method regards the light field decomposition as an overdetermined problem and uses the optimization algorithm to solve it.However,with the improvement of the resolution,depth,viewing angle,and other parameters of the reconstructed light field,the disadvantages of the optimization algorithms,such as low computational efficiency and large memory consumption,are magnified.It is difficult to realize fast computation.To address the issues,this paper proposes a new CLF decomposition algorithm.It decomposes the original light field into object points for storage which only occupies a small amount of memory;It uses depth weight and weighted average algorithm to replace the optimization algorithm of the CLF,which dramatically improves the calculation efficiency.Under the same conditions,the memory occupied by the proposed algorithm is only 38.8% of that of the traditional method,the calculation time is reduced by 93.7%,and the image quality is improved by about 1 dB.At last,the display effects of this method and the traditional method are compared through simulations and experiments,and the effectiveness of the algorithm is verified.

This paper proposes a multi-branch feature cascade image deraining network based on the attention mechanism to address the problems that existing deraining networks do not entirely deraining in diverse environments and do not adequately preserve image texture details.The model combines multiple attention mechanisms to form multi-branch networks to transfer and cascade the spatial image details and contextual feature information in the overall network and fuse them.Moreover,the stage attention fusion mechanism constructed between network branches can reduce the loss of image information during feature extraction and retain feature information to a greater extent,making the image deraining task more effective.The experimental results demonstrate that the new algorithm outperforms other comparison algorithms in terms of objective evaluation indices,the subjective visual effect can be effectively enhanced,the deraining ability is more substantial,the accuracy is more remarkable,and it can remove various densities of rain patterns while preserving the image's detail information.

To achieve high-precision detection of heart rate (HR) from face videos,this paper proposes a face video HR detection method based on the improved EfficientPhys network:TDM-EfficientPhys.First,the temporal derivative network (TDM-Net) is introduced to effectively preserve the temporal characteristics of the remote photoplethysmography (rPPG) signal.Then,the Dropout parameter of EfficientPhys is adjusted to embed the residual structure,which effectively avoids gradient disappearance and network degradation problems.In addition,TDM-Net and the improved EfficientPhys network are connected through skip connections to realize feature fusion of time series and spatial information to improve detection accuracy.In the end,two publicly available datasets,PURE and UBFC-rPPG,are used to validate the proposed model in this paper. Experiments show that root mean square error (RMSE) of the method tested in this paper is reduced to 2.387 and mean absolute error (MAE) is reduced to 2.040,which shows better network performance and more accurate HR compared with existing models.

An improved YOLO v5 model for cyclist helmet and license plate detection is proposed to solve the problems of low accuracy,poor generalization ability and single detection categories in helmet detection.Firstly,the convolutional block attention module (CBAM) is introduced into the backbone network to strengthen the key features of the target region and improve the accuracy of the model. Secondly,by optimizing the multi-scale feature module and adding a detection layer for tiny targets in the prediction end,the detection rate of the network for small targets in dense scenes is enhanced,and the generalization ability of the model is improved.Finally,the model training convergence speed is accelerated and target localization accuracy is improved by optimizing the bounding box regression using efficient intersection over union (EIoU) and by clustering new anchor box sizes using the K-means algorithm in the helmet and license plate dataset created.The experimental results show that the improved YOLO v5 model has achieved an increase in detection accuracy rate of 2.5%,a recall rate increase of 3.3%,and an average precision increase of 3.8%,which makes it more suitable for detecting helmet and license plate targets of cyclists.

Aiming at high mismatching rate in the traditional image matching algorithms and low measurement accuracy of binocular vision,a binocular vision measurement method based on nonlinear diffusion and high-dimensional modified-speeded up robust features (M-SURF) descriptor is proposed in this paper.Firstly,the nonlinear diffusion Perona-Malik (PM) model is improved to smooth the edge region and maintain the internal flat region unchanged in the image.Then,the diffusion image and the original image are differential operated to obtain the differential image,and the KAZE algorithm is used to detect the feature points.Secondly,the ring neighborhood is used to construct the descriptor.When the Harr wavelet response value is superimposed,the high-dimensional M-SURF descriptor is generated by multi-interval division according to the sign of the vertical response value;Finally,Hamming distance is used to match,and random sample consensus (RANSAC) algorithm is used to eliminate mis-matching and screen out the key matching point pairs required for measurement.The measurement can be completed by obtaining the 3D coordinates of the key matching point pairs according to the principle of parallel binocular vision measurement.The experimental results show that the matching accuracy of the proposed algorithm is 24.09% higher than that of the traditional KAZE algorithm,and the minimum relative error of measurement is 0.375 6%,which meets the requirements of measurement accuracy.

This paper proposes an improved dense residual U-shaped network (DRU-Net) to address the issues of blurry small vessel pixels and vessel discontinuity in retinal vessel segmentation.Firstly,the dense residual block (DRB) is proposed by combining the advantages of residual structure and dense connection,which is used to construct the encoding and decoding layers of the DRU-Net to fully extract the target features.Then,a multi-characteristic distillation module (MCDB) is added to the bottom of the network,which is built with dilated convolutions to extract image features at different scales.Finally,a bidirectional convolutional long short-term memory module (BConv LSTM) is introduced at the skip connection to fully fuse the shallow and deep features,and output the complete vessel map.Experimental results on the public datasets DRIVE and CHASE~~DB1 achieve an accuracy of 0.966 9 and 0.976 4,respectively.Meanwhile,the area under curve (AUC) reaches 0.983 9 and 0.986 7,respectively,which demonstrates that the network has good segmentation performance and certain application value.

In order to solve the difficulty of acquiring paired data in low-dose CT (LDCT) image denoising,a self-supervised LDCT image denoising algorithm based on attention mechanism and compound loss is proposed in this paper.In this algorithm, the feature extraction of LDCT images is completed by using the U-net network after edge enhancement. Channels and pixel attention mechanisms are introduced into the network framework to improve the ability of the network to suppress noise and artifacts.Moreover,in order to make the denoised images closer to the original images,we propose a self-supervised learning scheme with compound loss to avoid the over-smoothing phenomenon caused by the traditional loss.The experimental results show that the proposed algorithm can effectively suppress the noise of LDCT images and recover more texture details in LDCT images.The peak signal-to-noise ratio (PSNR) of the LDCT images processed by the proposed algorithm increased by 16.40% and the structural similarity (SSIM) increased by 9.60%.In the absence of paired data,the proposed method can effectively preserve the details and reduce the noise generated by low-dose scanning,which provides a new idea for clinical LDCT image denoising.

Due to the limitation of the existing retinal vessel segmentation methods that have insufficient segmentation ability for microvessels and capillaries,which leads to vessel disconnections and end vessel misses,resulting in poor retinal vessel segmentation performance,a multi-scale consistency and attention mechanism U-Net (MCAU-Net) is proposed.Firstly,the network embeds an attention refinement module (ARM) in the bottleneck feature layer,which can effectively refine the redundant features in the bottleneck layer and suppress the weights of irrelevant pixels,such as the background pixels.Moreover,the context fusion module (CFM) is combined with the traditional skip connection as a way to supplement the information gradually lost during the phase of feature extraction and strengthen the network′s ability to construct microvessels and capillaries.Finally,a multi-scale consistent training method is designed based on the multi-scale output of the network to enhance the sensitivity of the network to different scale features.The comparison experiments on DRIVE and CHASE~~DB1 public datasets show that the network in this paper has good segmentation performance.

In this work,a novel surface plasmon resonance (SPR)-based photonic crystal fiber (PCF) biosensor is investigated for real-time detection of severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2).By the interaction of the ligands on the sensing surface with the virus,the analyte′s refractive index (RI) ranges from 1.334—1.355.Three structures with different basic resonant materials are analyzed and evaluated based on the finite element method (FEM),including gold,silver,and silver with a titanium dioxide protective layer.The sensor′s wavelength works in visible,and the amplitude sensitivity (AS),wavelength sensitivity (WS),resolution,limit of detection (LoD) and figure of merit (FOM) reach up to -1 170.89 RIU-1,4 000 nm/RIU,2.5×10-5 RIU,6.3×10-9 RIU2/nm and 911.50 RIU-1,respectively.Results demonstrate that the proposed sensor has excellent potential in the detection of SARS-CoV-2,with excellent AS and LoD,which can be used for early and rapid detection.