Action recognition has always been an important task in the field of computer vision. There are mainly two tasks based on RGB video and human skeleton. The mainstream methods are 3D convolutional neural network and graph convolutional neural network. For the data modality of human skeleton, this work designs a graph convolutional neural network based on the self-attention mechanism. The algorithm can achieve advanced performance on skeleton-based action recognition tasks. In addition, a method is proposed to use deep supervision methods to supervise the intermediate features of video and human skeleton, which improves the coupling of the two data features and further improves network efficiency. The network structure of this algorithm is simple, and only 3.37×107 parameters are used to achieve an accuracy of 95.6% on the NTU-RGBD60 (CS) dataset.

Aiming at the problems of the current capsule endoscope, such as low resolution, limited field of view, and great influence by noise, a system solution was proposed. First, by introducing Q-type aspheric surface to correct aberrations, an endoscope imaging lens with a full field of view of 160°, a relative aperture of F#3.0, and a total system length of 4.3mm was finally obtained. MTF values at 140 lp/mm were all greater than 0.3. The imaging quality of an optical system depends not only on the performance of the lens, but also on the image sensor, especially in low-light environments. The noise model was obtained by analyzing the noise characteristics of each stage when the sensor was working. Using the established noise model, the capsule endoscope image data set was synthesized, and the neural network model was trained. The test results of the algorithm model show that the proposed comprehensive solution can effectively improve the imaging quality of capsule endoscope system.

Balanced homodyne detectors have the disadvantages such as small common mode rejection ratio (CMRR) and low response bandwidth, etc. To address the problems, we designed a scheme to test the response characteristics of the photodetector (PIN photodiode) in the detector and selected devices with similar performance for balanced detection to improve the CMRR. Furthermore, the scheme of cascading transimpedance amplification and proportional operational amplifier were proposed to extract the photocurrent signal, which solved the problem of the limitation of the gain bandwidth product of the primary sampling amplifier and improved the response bandwidth of the detector while ensuring the sensitivity. As a result, we achieved 300-MHz balanced homodyne detection and the CMRR went up to 66 dB. We also detected the quantum shot noise in the optical signal. These achievements provide an effective technical scheme for high-speed sensitive detection.

Action recognition is one of the basic tasks of computer vision. The skeleton sequence contains most of the action information, so skeleton-based action recognition has attracted a lot of research attention. Mathematically, the human skeleton is a natural graph, so graph convolution is widely used in action recognition. But ordinary graph convolution only aggregates low-order information between pairwise nodes, and cannot model high-order complex relationships between multiple nodes. To solve this problem, a multiscale hypergraph convolutional network is proposed, which aggregates richer information in the two dimensions of space and time, so as to improve the accuracy of action recognition. The multiscale hypergraph convolutional network has an encoder-decoder structure. The encoder uses the hypergraph convolution module to aggregate relevant information between multiple nodes in the hyperedge, and the decoder uses the hypergraph fusion module to restore the original skeleton structure. In addition, a multiscale temporal graph convolution model based on dilated convolution is designed, which is used to better aggregate the temporal-dimension motion information. The experimental results on NTU-RGB+D and Kinetics datasets verify the effectiveness of this algorithm.

Coherent Anti-Stokes Raman Scattering (CARS) is a stimulated Raman process, which has the non-resonant background (NRB), leading to peak shift and spectral distortion in the spectrum. In this letter, we use a femtosecond laser as the light source and a grating filter system generating narrow-band pump light. The femtosecond laser excites photonic crystal fiber, producing a supercontinuum spectrum as Stokes light. Two beams excite the samples simultaneously after modulated to circular polarized lights to produce CARS spectrums. We illustrate that circularly polarized light can effectively remove the non-resonant background in the CARS spectrum of anisotropic materials by simulation. Thus, the CARS spectrum has a similar spectral line shape to that of spontaneous Raman. The experimental results of CARS spectrums of polystyrene and liquid crystal samples generally agree with the calculations, proving that circularly polarized CARS spectroscopy is an effective method to remove the NRB of CARS spectrum.

The emission characteristics of fluorescent molecules near the surface of noble metals have changed significantly due to the influence of surface plasmon resonance (SPR). They are widely used in the design of nano-devices such as fluorescent probes. The energy transfer mechanism between fluorescent molecules and metals is the basis for the design of such fluorescent probes. Therefore, in this paper, surface energy transfer (SET) and metal regulated spontaneous emission effect of single fluorescent molecules in the Au/SiO2/Ag core-shell nanostructure were theoretically simulated by using finite difference time domain (FDTD) method. The SET between fluorescent molecules and the hybrid plasmonic mode was studied. The results showed that due to the local surface plasmon resonance coupling between gold core and silver shell, the energy transfer efficiency between fluorescent molecules and metals shows a d10 relationship, which d is the distance from the fluorescent molecules to gold surface. That result is obviously different from the conventional FRET effect and the SET effect of single metal structure. It is expected to be applied in the development of nano light sources and biosensors.

Quantifying the geometric structure of packing particles is of great value for understanding the macroscopic mechanical properties of particle systems. In this paper, Cell was utilized to measure the geometric structure of disk particles in a two-dimensional silo. The probability distribution of Cell shapes satisfied the exponential function distribution and was independent of the system size. Furthermore, it was observed that under the condition of vibration driving, the triangular-shaped Cell had a surge, and the formation probability increased by 19%, while the probability distribution of the rest shapes of Cell still satisfied the exponential function distribution. The experimental results reveal the structural characteristics of ordered packing in disordered particle system at intermediate scale, and provide references for perfecting the theory of particle mechanics.

In order to make the orbital angular momentum (OAM) mode generated in the optical fiber more integrated, a method to generate the orbital angular momentum mode using photonic crystal fiber (PCF) is proposed. By designing a special PCF, the transmission speeds of the eigenmodes of the odd and even modes are different, and thus a π/2 phase difference is formed after transmission for a certain distance to produce an OAM mode. The PCF is composed of a central air hole, a silica ring layer and an outer cladding layer. The outer cladding layer is composed of two layers of circular air holes. The number of circles in each layer is 2n+2, corresponding to the n-order OAM mode. The finite element method was used to carry out three-dimensional numerical simulation analysis of the designed optical fiber, and the +2, +3, +4, and +5 orders of OAM modes were successfully generated. The PCF method for generating OAM modes meets the trend of future integration and miniaturization for optical fiber communication systems, and has potential application prospects in all-fiber mode division multiplexing systems.

A 783 nm femtosecond fiber laser with long repetition rate was designed and constructed. The laser was based on the amplifying loop mirror mode-locked erbium-doped fiber oscillator and the environment-stable erbium-doped fiber double-stage amplifier cascaded by pulse separator. An average power of 1.3 W, pulse width of 130 fs, repetition rate of 77.1 MHz, and 1560 nm pulse output are achieved. By frequency doubling using periodically poled lithium-niobate , an average power of 0.5 W, a pulse width of 140 fs and a pulse output of 783 nm were obtained. Furthermore, the repetition frequency of erbium-doped fiber oscillator was traced to the rubidium atomic clock by repetition frequency monitoring and phase-locked loop technology. The peak value of frequency jitter is 5 mHz and the standard deviation is 1.2 mHz within 12 hours. The laser system has the characteristics of high stability, high integration and small volume.

Ghost imaging is an active indirect imaging technology based on second-order or high-order correlation of light field. It has the characteristics of high sensitivity and strong anti-interference ability, and has important application value in biomedical imaging, remote sensing imaging, ocean detection and other fields. However, the sampling time is long and the imaging speed is slow due to the requirement of sampling number in the process of ghost imaging. Achieving high quality image reconstruction under the condition of low sampling is still one of the problems that need to be solved in the practical application of ghost imaging. In addition to the detailed description of the principle of ghost imaging, this paper introduces the progress of low-sampling ghost imaging in recent years, such as the theoretical and experimental results of deep learning-based computational ghost imaging, and summarizes and prospects the application of ghost imaging.

Millimeter wave (MMW) images are widely used for the detection of contraband in human security screening. However, at this stage different pre-processing (PP) methods have a significant impact on the accuracy of MMW image target detection. To address this issue, we pre-processed the data by spatial filtering, spatial filtering plus maximum clustering and spatial filtering plus minimum clustering, and compared the accuracy of target detection by different methods. The experimental results show that all three methods can effectively improve the target detection accuracy, with spatial filtering being the best pre-processing method, achieving an accuracy of 92.3%, which is an improvement of about 4% compared to the original data without PP. It indicates that the spatial location-based filtering method is more effective than the reflection intensity-based clustering filtering method in improving the accuracy of MMW image detection, which can be used to improve the accuracy of the active MMW hologram. The task of improving the accuracy of active MMW hologram target detection provides a reference point.

A teaching and training system for infrared radiation characteristics measurement is presented. Based on the task of infrared radiation characteristic measurement equipment and the principle of equipment structure, nearly 20 operation subjects were designed. Firstly, the calibration principle of infrared radiation characteristic measurement system was described. Then, the software and hardware system of infrared radiation characteristic measurement teaching system were designed and implemented. The calculation model of infrared detector index test operation subject was studied. Finally, a series of experiments were carried out, and they showed that the system could achieve the requirements of teaching and training.