We propose a curvature sensing structure with low temperature sensitivity based on Mach-Zehnder interferometers (MZI),which is formed by large core-offset welding.As a core sensitive unit,GF3 fiber is welding with two sections of single mode fiber (SMF).The influence of core-offset and sensor length on temperature and curvature sensitivity is compared by experiments,respectively.Through these experiments,the optimal parameter combination to realize low temperature sensitivity curvature sensing is found.The experimental results show that when the sensor length is 10 mm and the core-offset is 20 μm,the temperature sensitivity of large core-offset MZI is only 0.9 pm/℃,and the curvature sensitivity can reach 10.9 nm/m-1 simultaneously.The sensor has the advantages of simple structure,low cost, high curvature sensitivity and low temperature sensitivity.It has potential application value in engineering structure,health monitoring and wearable device.

It is an interesting and practical problem to explore how to optimize the performance of quantum battery (QB) via quantum coherence.A QB model of a multi-level quantum system charged in a coherent auxiliary bath is constructed in this paper.Within the framework of collision model,the quantum master equation (QME) of QB charging in weak interaction regime is derived,and the solution of QME is obtained by the Ket-Bra entangled state method.Combining with numerical simulations,the influences of quantum coherence (the coherence magnitude and the coherence phases) on the maximal extractable work of QB are analyzed.It is revealed that quantum coherence,under certain conditions,can play the role of QB′s “fuel” to effectively improve the charging performance of QB.

In order to effectively prevent the excessive concentration of harmful gases from causing damage to economic benefits and the safety of human life and property,a new detection method has been proposed and designed.Through a series of tests on micro/nano ionizing sensors,the designed sensor exhibits strong bending resistance and fast response recovery ability.Multiple concentrations of benzene and toluene gases are tested under experimental conditions of room temperature,atmospheric pressure,and relative humidity of 75%.Corresponding recognition models are constructed based on the intensity of characteristic noise to identify the concentration of the gas.The experimental results show that the fitting degree between the measured value and the true value of the sensor for detecting harmful gases can reach over 99%,reflecting high detection accuracy.Moreover,the sensor is in a reversible ionization equilibrium state when detecting harmful gases,with good repeatability and without preheating.It is safe and non-toxic.Therefore,it is suitable for application in the detection of benzene gases,meeting the requirements of efficiency and accuracy,and has certain practicality.

Aiming at the issue of low accuracy of rainfall estimation by the machine vision algorithm,a rainfall inversion algorithm based on social surveillance video is proposed.Firstly,the rainfall classification network is adopted to remove the no-rain video.Secondly,the foreground information of rainfall videos is extracted by using the alternating direction method of multipliers (ADMM),and the region of interest (ROI) is chosen by semantic segmentation and background subtraction methods.Thirdly,a Gaussian mixture model (GMM) characterized by gray-scale change and saturation features is constructed to choose the raindrops in ROI.Finally,the raindrop size is calculated according to the perspective imaging relations,and the rainfall is inverted through the meteorological Gamma model.The experimental results show that the rainfall classification accuracy of the method reaches 91.3% in the multi-class weather dataset (MWD) and 77.0% in the real dataset,and the rainfall estimation results are more accurate compared with the existing methods.

A lightweight DGC~~YOLOv5 (you only look once v5) algorithm is proposed to solve the problems of poor detection capability of small targets,large size of the model,complex calculation,and difficult deployment on mobile devices for flame detection model.Firstly,the k-means calculation function is used to calculate the anchor size for this data set.Secondly,the convolutional block attention module (CBAM) is introduced to improve the detection ability of this algorithm to small target.Then the lightweight Ghost module is adopted to improve the C3 modules in backbone network.Finally,the depthwise separable convolution (DS~~Conv) which uses simple linear calculation instead of complicated calculation is used to reduce model complexity and size.Experiments show that compared with the original YOLOv5 algorithm,the mean average precision (mAP) of the proposed algorithm can reach 94.4% on the test set,1.7% higher than the original algorithm.The average detection speed of the proposed algorithm can reach 71 FPS on the video test set,which can meet the requirements of real-time detection.Parameters and the floating-point operations (FLOPs) calculating amount are respectively reduced to 41.2% and 34.8% of the original algorithm,and the model size is reduced by 8.4 M,which facilitates the subsequent deployment on mobile devices.

In natural scene text recognition,a fixed size convolution kernel is used to extract visual features,and then character classification is performed.The global modeling ability of this method is weak and it ignores the importance of text semantic modeling.Therefore,this paper proposes a natural scene text recognition method based on character attention.Firstly,a multi-level efficient Swin Transformer network is constructed to extract features,which is different from the convolutional network.This network can make the features of different windows interact with each other.Secondly,the character attention module (CAM) is designed to make the network focus on the features of the character region,so as to extract the visual features with higher recognition ability.Then,the semantic reasoning module (SRM) is designed to model the text sequence according to the context information of characters.And the module can obtain semantic features to correct the indistinguishable or fuzzy characters.At last,visual and semantic features are fused to get the results of character recognition.The experimental results show that the recognition accuracy in this paper reaches 95.2% on the regular text data set IC13 and 85.8% on the irregular curved text data set CUTE.The feasibility of the proposed method is proved by ablative and comparative experiments.

Aiming at overcoming the limitations of under enhancement,over enhancement and low contrast in the existing image enhancement methods,a low light image enhancement method based on guided filter and pixel redistribution is proposed.This method estimates the illumination part of image with guided filter,taking full advantage of the edge preservation capability of the guided filter;the pixels of the illumination image are subjected to relatively uniformly redistribution, so as to comprehensively improve the brightness and contrast of illumination image.The final enhanced image is obtained by performing inverse Retinex transformation on the illumination image enhanced by pixel redistribution and the refleced image.The experimental results confirm that compared with existing enhancement methods,the proposed method achieves better enhanced effect,the contrast and textural structures of enhanced image are clearer.

The realization of multi-armored target tracking plays a vital role in cooperative tracking and strike,and the realization of multi-armored target tracking needs to solve the problems of occlusion,interspersed and constantly changing target scales between targets.Therefore,an online multi-armored target tracking method based on visual-attention Gabor filter is proposed to achieve the tracking of multi-armored targets in the ground battlefield.A visual-attention Gabor filter branch is constructed to enhance detection by simulating the retinal structure.By introducing temporal information,the problem of target occlusion is solved by using an online learned target-specific convolutional neural network.What is more important,a multi-armored target tracking dataset is constructed by means of actual shooting and downloading from the internet,and the current mature multi-target tracking methods are compared with the method proposed in this paper through experiment.The experiments show that the method in this paper not only has excellent tracking performance,but also can meet the actual application requirements.

Aiming at the specific situation of the fuzzy and noisy in the printed circuit board (PCB) photoelectric image,an improved edge information extraction algorithm is presented.Firstly,the adaptive fuzzy set enhancement algorithm and the edge detection algorithm of mathematical morphology (EDAMM) are improved,respectively,and their basic principles are analyzed.Then the two algorithms are combined to preprocess the PCB photoelectric images and extract their edge information.Finally,the experiments for edge information extraction are carried out with two PCB photoelectric images acquired by different imaging systems.The results show that the contrasts between light and dark of two images obtained by this algorithm are higher,and the accurate and clear edge information is extracted,the noise is significantly reduced.And the excellent quality coefficient of the obtained image is higher,which are 0.885 2 and 0.874 9 of the two images,there are higher than which of the other four algorithms mentioned in this paper,respectively.This shows that our algorithm can better remove the fuzziness and noise of PCB photoelectric images,and it can accurately extract their edge information.

Solar cell is the core component of photovoltaic power generation system,their optical reflection characteristics have an important theoretical significance for the research of power generation efficiency and defect detection of PV module,but it is difficult to observe and measure them directly in practical applications.In this paper,the optical transmission process of PV module is studied based on Fresnel′s law and multilayer media model.Firstly,the bidirectional reflection transmission model of solar cell surface is established based on the microfacet theory.Then the polarization bidirectional reflection distribution function (BRDF) of the solar panel is derived with the help of multilayer media model,the optical polarization characteristics observation platform of PV module is built by polarization camera and polarization experiments are carried out.Based on the above mentioned model, the genetic algorithm is used to invert the model parameter from the experimental measurement data and the simulation curve of polarization information with the observation angle is derived.The results show that the experimental observation data of the optical transmission model proposed in this paper can be in good agreement with the simulated values,which provides a new theoretical reference for studying the defect detection of solar panel surfaces.

In order to solve the Lambert model is difficult to calculate the indoor visible light channel noise and error problems,a neural network algorithm is proposed to realize the indoor visible light channel model.Aiming at the problems of large amount of fingerprint database data,difficult collection and many training parameters,which lead to slow iteration speed,the generative adversarial network (GAN) is proposed to generate simulation data set and merge the original sparse fingerprint database to generate the number of fingerprint database meeting the training requirements.A one-dimensional convolutional neural network (CNN) is used to extract data features,reduce training parameters and improve iteration speed.The sparse fingerprint database is collected in the indoor environment of 5 m×5 m×3 m,and the back propagation neural network (BPNN) and one-dimensional CNN indoor visible light channel model are respectively used for comparison.The simulation results show that the average absolute error of GAN is 0.04,and the data volume is increased by 300%.Under the same fingerprint database,the error of BPNN channel model is 3.81,and the convergence is realized after 500 iterations.However,the error of CNN channel model is 0.79,and the iteration converges are 100 times.The GAN fingerprint database merged CNN visible light channel model proposed in this paper has the advantages of high precision,small error,fast speed and strong generalization,which provides a new research scheme for indoor visible light channel model.

In order to solve the imbalance between watermark hiding degree and robustness in the watermarking algorithm,the quick response (QR) code and holographic technology are combined in the Contourlet-SVD domain.A digital holographic watermarking algorithm based on QR code in the Contourlet-SVD domain is proposed.First,the different carrier images are chosen to perform 3-layer Contourlet transform.In view of the advantage of low frequency coefficient,the singular value decomposition (SVD) joint transform is performed for the low frequency coefficient.Second,the QR code which is encoded by “Flight School” image is chosen as the original watermark image.The QR code holographic watermark is obtained by conjugating symmetry extended Fourier digital hologram.Thirdly,the holographic watermark is superimposed on the singular value obtained by SVD to complete the embedding of the watermark information.The simulation experiment proves that the image clarity of the algorithm is better after embedding the watermark information.The peak signal-to-noise ratio (PSNR) value of the algorithm is above 31 dB.The similarity normalized correlation (NC) of the algorithm reaches 0.989 0.The extracted watermark information can be clearly identified after the image has been tested against resistance.The NC values all reach above 0.90 especially for salt and pepper noise,Gaussian noise and mid-pass filtering attacks.

Accurate segmentation of pathological cell nucleus is the basis of pathological diagnosis,however,the current algorithm for automatic segmentation of mitotic breast cancer cell nucleus is poor.This paper analyze and study the current cell nucleus segmentation algorithm.And a U-shaped network (U~~net) based on a combination of attention mechanism and residual structure is proposed to solve the problem of insufficient accuracy of cell nucleus segmentation due to the close morphological characteristics of mitotic and non-mitotic cells.The experiments show that the average pixel accuracy (MPA) and Mean~~dice index coefficient of the proposed algorithm are 0.74 and 0.82. Compared with the original algorithm, the training indexes are improved by 11% and 9%,which proves the feasibility of the algorithm in this article.

In order to solve the problem that the fusion algorithm of low-rank decomposition and sparse representation (SR) causes a lot of information missing,a brain image fusion algorithm combining latent low-rank decomposition and SR is proposed.Firstly,the source image is decomposed into low-rank,sparse and noisy components.In the face of the differences between the characteristics of different decomposition components,the low-rank and sparse dictionaries are constructed to describe the low-rank components respectively.The weighted gray value method is used to process low-rank components to maintain their contour and brightness features.For the sparse components,a multi-norm weighted metric method is designed to improve the SR to maintain the high-dimensional information.The noise components are eliminated.Compared with the current five mainstream algorithms,the proposed method has the best effect in terms of visual effects and objective indicators.