The power tuning characteristics of laser diode end pumping dual-wavelength laser with combined gain medium are studied.The results show that the waist position of pump beam in combined gain medium and the temperature of gain medium are important factors affecting the relative power ratio of the dual-wavelength signals.When the pump beam waist position is fixed,the different change rate of the emission cross section of gain medium will cause the relative power of the dual-wavelength signal to change with temperature.The experiments show,for the combined gain medium dual-wavelength laser with specific parameters,when the pump power is 3 W,and the pump beam waist position is 1.5 mm in the gain medium,the heat sink temperature rises from 5 ℃ to 32.3 ℃,the output dual-wavelength signal achieves power balance,and the total output power is 435 mW.The experimental results are in good agreement with the theoretical simulation.

Silver is a potential candidate of flexible transparent electrode materials due to its excellent conductivity,malleability and ease preparation.Ag thin films with different thicknesses (6 nm,10 nm,14 nm,18 nm,20 nm,24 nm) were prepared by vacuum thermal evaporation.Because of the combined effect of light scattering and absorption,the transmittance decreased first,then increased,and then decreased with the increase of thickness.When the thickness of Ag film was 18 nm,the best transmittance was about 60%.Meanwhile,the surface resistance decreased gradually with the increase of thickness.In order to improve the transmittance of Ag film,molybdenum trioxide (MoO3) with high refractive index of 2.1 was introduced to modify Ag film to prepare MoO3/Ag/MoO3 (MAM) multilayer film.The results show that the introduction of MoO3 smoothes the surface of Ag film,reduces the surface resistance and improves the conductivity.More importantly,the refractive index coupling occurs at the MoO3/Ag interface,which greatly improves the overall transmittance of the multilayer,and the transmittance generally increases by at least 10%.When the thickness of Ag is 14 nm,the transmittance of MAM multilayer is the best,up to 70 %.Finally,the feasibility of silver transparent electrode was verified in a green organic light-emitting diode (OLED).

A modulated grating Y-branch laser (MGY) is an ideal light source for fiber grating demodulation system.At present,there are few modeling studies on its tuning characteristics.In order to study the tuning characteristics of MGY,this paper extended the transmission line laser model (TLLM) to simulate MGY.The Y-type laser with analog modulated grating is studied by combining the TLLM,transmission matrix method (TMM) and digital filtering method.In the combined model,we use the TMM method based on time domain for the gain region and phase region.Firstly,the TMM based on frequency domain is used for the left and right modulated grating (LMG/RMG) regions,and then the inverse Fourier transform is used as the time domain model.The output current characteristic curve and static tuning characteristic of the MGY are simulated successfully,which is similar to the published experimental results.This comprehensive model can be used to study the transient response and laser spectra of devices that are difficult to obtain using frequency-domain models.

Based on the principle of inter-mode interference,a Mach-Zehnder interferometer (MZI) sensor composed of Panda polarization maintaining fiber (PMF) is fabricated in this experiment.Because of the coupling of the large aperture multimode fiber (MMF),the sensor shows high sensitivity to temperature.When the external temperature changing,the transmission spectrum of the sensor shifts.By observing the wavelength shift of characteristic peaks,the temperature response characteristics of the sensor is obtained.From the experimental data,the wavelength of two characteristic peaks of the SMF-MMF-PMF-SMF interferometer structure linearly responses to temperature.The sensitivity of temperature is －123.80 pm/℃ and －195.20 pm/℃,respectively.The repeatability and stability of this sensor for temperature measuring are very good,and this sensor can effectively measure the ambient temperature.

The proposal of MonoDepth2 has made significant progress in self-supervised monocular depth estimation,but the prediction effect of the network in large non semantic regions and boundaries is not ideal.The main reason is that the basic U-Net framework does not make full use of multi-scale feature information,resulting in poor depth estimation from large gradient regions.To address this problem,this paper proposed an improved DepthNet,a hierarchical integration net (HINet).The U-Net network structure is optimized so that the encoder side can generate feature information of different scales at each layer,thus allowing the decoder side to fully fuse multi-scale features at each layer.Since the feature information of different scales contributes to a specific decoder layer to different degrees,the hierarchical integration (HINet) algorithm proposed in this paper also adds a channel attention module to enhance the weight of important feature scales.When stereo pairs are used for training,this paper preprocesses the data and adds a depth-implying loss function for stereo pairs.The experimental results on the KITTI dataset show that all indicators are improved to varying degrees,in which the absolute relative error is reduced by 0.09 and the squared relative error is reduced by 0.093.

To further improve the image segmentation accuracy,improving the traditional multi-threshold image segmentation method with large computation and slow segmentation,we proposed a multi-threshold image segmentation scheme.First,the initial solution is optimized by using the cubic chaotic mapping to improve the search efficiency.Then,scaling factors of the eagle perching optimizer (EPO) and crazy operators are introduced for perturbation and combined with position updates of the sparrow search algorithm (SSA),to improve the optimization accuracy,convergence rate and avoiding the local optimum.The improved seagull optimization algorithm (ISOA) is tested for performance using six benchmark functions.Finally,the ISOA is combined with threshold optimal selection for multi-threshold image segmentation based on Otsu and compared with existing segmentation algorithms.Simulation results show that the ISOA achieves the optimal value for 100% of the Otsu-based segmentation,and 80.9% outperforms the rest,optimizing both the segmentation accuracy and quality of the image.

Based on tracking algorithm of learning discriminative model prediction for tracking (DIMP),a discriminative single target pedestrian tracking algorithm with adaptive tracking state is proposed to address the problems of unstable tracking state due to background similarities interference,mutual occlusion between pedestrians and background cluter encountered in the pedestrian tracking process.The response map is obtained by the convolution operation of the classification filter and the search region in the tracking process,and the tracking state is divided into weak response state,multi-peak strong response state,and single-peak strong response state by the response map.For the influence of disturbances in the multi-peak strong response state,an online update strategy is proposed to update the classification filter by using the excitation and suppression losses to improve the discriminative ability of the classification filter.For the problem of inaccurate target prediction in multi-peak strong response and weak response states,the target position is corrected by offset and adding candidate frames to improve the tracking accuracy. The proposed algorithm is experimentally verified, which achieves precision of 0.978 and a success rate of 0.740 on pedestrian video sequences with a real-time speed of 30 fps under NVIDIA GTX 1650.

Deep learning technology is widely used in target detection tasks because of its powerful feature extraction capabilities.Aiming at the problems of uneven recognition accuracy and low detection efficiency of multi-scale cervical cancer cells,this paper proposes an improved recognition algorithm,mini-object-YOLO v3 (mo-YOLO v3) based on the YOLO v3 model.The cervical cell images collected under a 20× digital scanner are selected as the data set.In order to improve the robustness of the algorithm,multiple data enhancement strategies such as contrast enhancement,grayscale image,rotation and flipping are introduced to expand the data set;the model takes Darknet53 network combined with attention mechanism as the backbone module,for the large difference in the size of cervical cancer cells,a multi-scale feature fusion algorithm is proposed to optimize the model structure.In order to solve the problem of low detection accuracy of small targets,an improved loss function is proposed,adopting the relative position information method to reduce the influence of the object frame on the detection result.The test results show that the mo-YOLO v3 model proposed in this paper not only has obvious advantages in overall recognition accuracy,but also greatly improves the positioning accuracy of small-size cervical cancer cells.The model has an accuracy rate of 90.42% for identification of cervical cancer cells,a precision rate of 96.20%,a recall rate of 93.77%,and a similarity index ZSI of 94.97%,which is higher than similar algorithms.

Aiming at the problem that it is difficult to effectively extract the key information of pedestrians in the chaotic scene and the global feature method is invalid in the case of partial occlusion,a multi-granularity person re-identification (ReID) method guided by a double pyramid structure is proposed.First,the attention pyramid in is embedded ResNet50 to guide the network to dig out features of different granularities from coarse to fine,making the network more inclined to focus on the significant areas of pedestrians in complex environments;secondly,the branch of the double attention feature pyramid (DFP) with asymmetric structure is adopted.Multi-scale pedestrian features are extracted to enrich the diversity of features.At the same time,the dual attention mechanism allows branches to capture finer-grained local features from shallow information;finally,the coarser-grained global features are merged with multi-level and fine-grained local features,The two kinds of pyramids interact to retain more discriminative multi-granularity features to improve the pedestrian occlusion problem.Experiments on multiple data sets have shown that the evaluation indicators are higher than most current mainstream models.Among them,on the DukeMTMC-reID data set,Rank-1,mAP and mean inverse negative penalty (mINP) reached 91.6%,81.9% and 48.1%,respectively.

The recognition of oral mucosal diseases mainly depends on doctors′ visual observation and subjective judgment.This method leads to low accuracy of disease recognition and heavy workload of doctors.To solve the above problems,an oral mucosal disease recognition method based on multi-level feature fusion is proposed.There are two kinds of deep-level features and shallow features extracted from oral disease images.The efficientNet model is used to extract the deep features.HSV,histogram of oriented gradiant (HOG) and gray level co-occurrence matrix (GLCM) are used to extract the shallow features of color,shape and texture of oral diseases respectively.After feature fusion,the random forest (RF) algorithm is used to select the features with greater feature importance,reducing the dimension of the feature.Finally,a variety of machine learning classifiers are combined for classification and recognition.The datasets of oral mucoal diseases collected are used for experiment verification.The experimental results show that the method can achieve the accuracy (Acc) of 92.89%,sensitivity (Sen) of 89.91%,specificity (Spe) of 96.06% and area under the curve (AUC) of 98.09%.It effectively solves the problems of many misjudgments and low accuracy in recognition.

In order to solve the problem of low accuracy of person re-identification (ReID) algorithm in complex environment caused by unclear detail features and changeable attitudes of pedestrians,a ReID network based on multi-granularity feature extraction and feature fusion is proposed.Firstly,two granularity partitioning methods are used to obtain the local features of the image at the input and output ends of the backbone network.Secondly,spatial transformation network (STN) is introduced to align the global image and enhance the local image.Finally,local feature fusion is used to mine the correlation information between features to improve the model′s ability to recognize similar samples.Experimental results show that the proposed method achieves good recognition performance on multiple datasets.The mean average precision (mAP) and first accuracy (Rank-1) of the market-1501 dataset are 84.87% and 94.45%,respectively.Compared with the current mainstream ReID algorithms,the proposed method has better recognition effect.

Because pedestrians are easily affected by background,occlusion,posture and other issues in real scenes,in order to obtain more discriminative features in pedestrian images,a person re-identification method based on attention mechanism and local association feature is proposed.Firstly,the attention module is embedded in the network framework to pay attention to the features with strong expressive ability in the image.Then,the local association features are obtained by using the association of adjacent parts in the image,and combined with the global features.The experiments on Market1501 and DukeMTMC-ReID datasets show that the Rank-1 index reaches 95.3%and 90.1%,respectively.The results show that the proposed method can fully obtain the feature information with strong discrimination and make the model have strong recognition ability.

Video summarization is an effective way to quickly obtain the key information of video.Existing video summarization methods usually have high computational complexity and are difficult to be applied in the scene with limited computing resources.Therefore,a spatiotemporal joint method for surveillance video summarization via the direction information is proposed.This method first uses the horizontal slice to obtain the object spatiotemporal motion trajectory.Secondly,the spatiotemporal trajectory background is eliminated and the linear trajectory slope is calculated,and the object motion direction is determined according to the object spatiotemporal trajectory slope. Thirdly,the motion segment in the sampling domain is detected to determine the timing position of the object in the video.Finally,the video summarization is constructed adaptively according to the object timing position and motion direction.Experimental results show that the average frame processing time (AFPT) of the proposed method reaches 0.374 s,which has obvious advantages over those of the comparison methods.In addition,the generated video summarization is concise and efficient,and the user experience is good.

Based on the generalized Huygens-Fresnel principle,the analytical expression of the cross-spectral density function for the partially coherent Laguerre-Gaussian (PCLG) beam propagating in biological tissues is derived,and used to study the effect of propagation distance 〖ＷＴＢＸ〗z,topological charge m,radial index n,and spatial coherence length σ0 on the normalized intensity and phase evolution of the beam in the dermis of mouse tissue.It is shown that as the propagation distance increases,the intensity gradually decreases,the intensity distribution profile slowly evolves from multi-peak shape to single-peak shape.Initial m-order coherence vortex and n circular edge dislocations evolve into m+2n coherent vortices firstly,and then m+2n coherent vortices are newly created.In addition,the larger the topological charge,radial index,and spatial coherence length,the slower the change in the intensity distribution and phase evolution.