In order to improve the ambient humidity sensitivity of fiber Bragg grating (FBG), coating graphene oxide (GO) polyvinyl alcohol (PVA) solution on FBG for humidity sensitization is proposed in the paper. Firstly, the humidity response of the sensor is studied theoretically, and the sensor samples are fabricated by pulling coating method, then the influence of the humidity sensitive solution mixture ratio and the film thickness on the humidity sensitivity is discussed experimentally. Further experimental results show that when the mixture ratio of graphene oxide-polyvinyl alcohol is 9∶1 and the coating thickness is 89 m, the humidity sensitivity of the sensor is 2.3 pm/%RH (relative humidty) in the humidity range of 50%—60%RH. The humidity sensitivity of the sensor is 7.7 pm/%RH in the humidity range of 60%—70%RH. The humidity sensitivity is 47.2 pm /%RH in the humidity range of 70%—80%RH. The teumperature sensitivity is 13.2 pm/℃, humidity measurement instability is 0.064 7%RH, and response and recovery time is 0.84 s and 11.46 s, respectively. The sensor has advantages of simple manufacture, high humidity sensitivity and good stability, which makes it has great application potential.

Polarization converter is widely used in many engineering fields such as wireless communication, radar detection and so on, aiming at the current application requirements of ultra-wideband, multi-function and high-efficiency, a dual-polarized, multi-band reflective polarization converter is proposed. Based on the metasurface, the structure is composed of three parts: metallic patch arranged periodically in top layer, dielectric layer and metallic plate in bottom layer. The research shows that for the incident linear polarization wave, the structure can realize cross-polarization conversion in the three bands of 6.2—6.7 GHz, 17.7—19.2 GHz, 20.88—21.14 GHz with a polarization conversion ratio (PCR) of over 90%. It can realize linear-circular polarization conversion in the two bands of 7.12—16.68 GHz, 19.47—20.67 GHz with a axial ratio (AR) of below 3 dB. This design has the characteristics of simple structure, convenient manufacture and low-energy consumption, it has certain theoretical and practical value in electromagnetic wave polarization controlling and microwave device designing and so on.

To address the issues of poor robustness and large model parameters in existing stereo matching algorithms in areas such as weak texture images, the PSMNet stereo matching method is improved by using an atrous spatial convolutional pooling pyramid structure (ASPP) to extract spatial feature information of images at different scales. Subsequently, a channel attention mechanism is introduced to assign corresponding weights to feature information at different scales. The above information is integrated to construct a matching cost volume, an hourglass shaped encoding and decoding network is used to standardize it, and determine the correspondence between feature points in various disparity situations. Finally, the linear regression is used to obtain the corresponding disparity map. Compared with PSMNet, the error rates of this study in the SceneFlow and KITTI2015 datasets are reduced by 14.6% and 11.1% respectively, and the computational complexity is reduced by 55%. Compared with traditional algorithms, it can improve the accuracy of disparity maps and enhance the quality of 3D reconstructed point cloud data.

In the task of vehicle re-identification (Re-ID), joint extraction of global and local information has become the mainstream approach. However, many Re-ID models only focus on the richness of local information while neglecting completeness. To address this issue, an algorithm based on relationship fusion and feature decomposition is proposed in this paper. The algorithm starts from the spatial and channel dimensions, dividing the features extracted by the backbone network along the vertical, horizontal, and channel dimensions. Firstly, to better highlight the foreground region of the vehicle, a mixed attention module (MAM) is proposed. Then, to explore rich feature information in the spatial dimension while making the network pay attention to more complete regions of interest, graph-based relation fusion is designed for the segmented features in the vertical and horizontal directions. To endow the network with the ability to capture more discriminative information, feature decomposition is implemented on the segmented local features in the channel direction. Finally, vehicle Re-ID is achieved through the joint effect of the features extracted from the global branch and the robust features from the local branches. Experimental results demonstrate that the proposed algorithm achieves state-of-the-art performance on two popular vehicle Re-ID datasets.

In order to solve the problems of small number of training samples and complex background of remote sensing images, this paper introduces transfer learning and channel attention into convolutional neural network (CNN), and proposes a remote sensing image scene classification method based on transfer learning and channel attention. Firstly, this method selects two CNNs pre-trained by ImageNet natural dataset as the backbone, and introduces the channel attention mechanism to adaptively enhance the main features and suppress the secondary features. Then the features extracted from these two networks are fused for classification. Finally, fine-tuning transfer learning is used to realize learning and classification in the target domain. The proposed method is evaluated on several classical public datasets, and the experimental results show that the proposed method achieves the same performance as other advanced methods in remote sensing image scene classification.

Aiming at the problems of large number of parameters of YOLOv5s model and difficulty in deploying on embedded devices, a lightweight YOLOv5s strip surface defect detection method is designed. Firstly, part of the convolutional layer in the backbone network is replaced with RepGhost with multi-branching structure, which enhances the ability of the backbone to extract feature information, and the reasoning can be converted into a single-branch structure to ensure the detection speed. Secondly, a lightweight FPN network (GG-FPN) is proposed, in which G-Ghost is used to reduce redundant parameters in the C3 module, while GSConv uses the large convolutional kernels depth separate convolution and branching structures to ensure the improvement of accuracy and speed at the same time. Experiments show that on the NEU-DET dataset, the number of parameters of the GG-FPN model is reduced by 24.7% compared with the original FPN, and the GFLOPs are reduced by 20.6%. For the whole model, the improved algorithm mAP only loses 1.9%, the number of parameters is reduced by 37.5% compared with YOLOv5s, GFLOPs is reduced by 33.1%, and the detection speed reaches 187 frame/s, which better balances the speed and accuracy of detection.

Surface plasmon resonance (SPR) has tremendous application value in biosensing. The phase sensing of the resonant angle based on a prism structure exhibits higher sensitivity compared with the reflectivity sensing. Multi-channel real-time phase shift detection is a research highlight in SPR biosensing. An SPR phase detection system based on Kretschmann prism structure attenuated total reflection (ATR) single- path is proposed. The phase difference between P-polarized signal light and S-polarized reference light is controlled by a liquid crystal variable retarder (LCVR) to realize single-path interference. Fixed-point phase detection of two-dimensional surfaces is realized by an addressable micro-mirror array (MMA). The system can detect the phase shift information with a spatial resolution close to a single MMA pixel (7.64 m horizontally/vertically or 10.8 m diagonally). It is capable of realizing single-path, high-precision, multi-channel two-dimensional plasmon resonance phase shift detection, providing new experimental ideas for in-situ biosensors and phase imaging beyond diffraction limits.

It is always been the difficulty to measure flow rate and phase content ratio of oil-water two-phase flow for the petroleum development. In this paper, the distributed acoustic sensing (DAS) technology is proposed for the acquisition and transmission of oil-water two-phase flow vibration signals of pipe wall. Axial stretching of the winding fiber consists two components: turbulent eddies vibration and sound speed of the turbulent flow. Then the mixing flow rate and phase content ratio are calculated by the fiber phase amplitude. To improve the sensitivity of vibration signal measurement with fiber, a winding fiber sensing model is proposed. We analyze the effects of two parameters, winding pitch and the number of tightly winding turns, on the vibration amplitude of the pipe wall and the sound speed along the axial direction of winding fiber, and experimentally verify three of the winding methods. It turns to more accurately measure the vibration of the pipe wall and the sound speed of the fluid with the winding method of 4 tightly winding turns at a pitch of 0.02 m. The R2 of fitting curves between the standard deviation of the fiber phase amplitude and the flow velosity within the pipe reaches 94%, and the calculated relationship between sound speed and oil content ratio is fully consistent with the theory.

This paper proposes a dual enhanced network (DEN) based on channel and feature learning to address the problem of poor accuracy in cross- modality person re-identification (ReID) caused by heterogeneous sample differences, person occlusion, and background interference. At the channel level, visible channels are randomly swapped to explore the relationship between visible and infrared channels, enhancing the model's robustness to multimodal sample changes. At the feature level, a normalization-based attention module (NAM) is introduced before module sharing network to avoid noise interference on modality-invariant information learning by punishing weights with smaller contribution factors, and a feature separation module (FSM) is used to separate identity-related features from identity-independent features, improving the model's recognition ability for heterogeneous samples. Finally, the network is trained and supervised using hard sample triples and weighted regularization loss to constrain pedestrian feature learning. On the RegDB dataset, DEN achieves a high level of accuracy, with a Rank1 accuracy of 94.86% and mAP of 90.10%.

To solve the problem that PointNet++, a direct point cloud data processing deep neural network, cannot thoroughly learn the shape information of point cloud, and SAFE-PointNet++ (spatial awareness and feature enhancement PointNet++), a 3D point cloud classification and segmentation method is proposed, which combines both spatial awareness module and feature enhancement module (SAFE). Firstly, the spatial awareness (SA) module is designed to help the feature extraction network integrate the weight information of spatial structure when the feature dimension is raised, thus enhancing the expression function of the feature in space. Secondly, the feature enhancement (FE) module is designed so that the additional information of the point cloud can be fully used by respectively splitting and encoding the enhanced geometric information and additional information. The experiment results show that SAFE-PointNet++ achieves higher classification and segmentation accuracy than the other ten classical networks on ModelNet40 and S3DIS datasets.

The exponentially increasing pending information volumes in the era of artificial intelligence require digital computer with higher energy efficiency. Thus, it is urgent to develop new computer technologies to address this challenge in the post Moore era. Inspired by the parallel brain's computing architecture, synaptic analogues based on the integrated storage and computing architecture is considered as an effective technology to solve the traditional von Neumann bottleneck. Our article highlights the memristive mechanism of neuromorphic devices, and also categorizes and interprets recent research works on synaptic devices according to programming modes. Besides, based on device properties makes a introduction of their potential applications in perceptual simulation or other aspects. Finally, the key technical challenges faced by current neuromorphic devices are summarized, and their future development prospects are prospected.