Abstract:A crack detection network based on improved YOLOX(Crack-YOLOX)is proposed to address the problems of poor accuracy and low efficiency of existing crack detection networks for highway pavements.The network's ability to sense pave-ment cracks is enhanced by introducing a simple parameter-free attention mechanism,a weighted bidirectional feature pyramid net-work is used to enhance the network's ability to represent small-size pavement crack features,and the HardSwish activation func-tion is used to speed up the network computation.Eiou and varifocalloss loss functions are introduced to improve the network regres-sion accuracy.Tests are conducted on a publicly available dataset,and the test results show that the network outperforms the com-parison network in terms of mean average accuracy and frame rate per second transmission.

Abstract:Synthetic aperture radar(SAR)is one of the main means to implement large-scale marine monitoring.In order to reduce the detection range,it is necessary to segment the SAR image from sea to land.According to the gray level characteristics of SAR images,this paper proposes a SLIC super pixel segmentation method for SAR images based on the adaptive threshold of image complexity.The variance between pixels is used as the judgment basis for the super pixel cluster center,and the variance threshold is adaptively adjusted according to the image complexity to reduce the number of sub domains for gradient calculation as much as possible.The results show that,compared with the traditional super pixel segmentation method,the algorithm based on the adaptive threshold of image complexity is significantly faster.

Abstract:This research introduces a novel type of brain tumor segmentation network called NFGAN that handles the difficul-ties of model design and sample category imbalance.The generator model based on the full convolutional network can actually real-ize the segmentation network's end-to-end output and enrich the semantic information of brain tumors through lateral connections.NFGAN can also strengthen the essential characteristics of the learning data through confrontation training in order to improve the segmentation accuracy of brain tumors.This paper also proposes a novel loss function that is a linear combination of weighted cross entropy loss and generalized dice loss in order to reduce the influence of category imbalance.On the BraTs2018 dataset,the pro-posed model shows segmentation accuracies of 87.9%,81.2%and 77.8%,with superior segmentation outcomes than the classic U-Net model.

Abstract:Rumors affect all kinds of human affairs,may destroy the market order,lead to serious social chaos.In order to ac-curately understand the propagation process of rumors from the micro mechanism,effectively reduce the harm of rumors to the soci-ety,in this paper,a SIHR rumor propagation model based on group propagation is constructed on a scale-free network,the corre-sponding average field equation is derived,and the steady state analysis is carried out.The next generation matrix method is used to calculate the basic regeneration number of the model in a scale-free network.Monte Carlo method is used to numerically simulate the SIHR rumor propagation model based on group propagation in BA scale-free network and WS small-world network.The simula-tion results show that under the same parameters,the rumor propagation time of this model in BA scale-free network is less than that in WS small-world network.Group propagation speeds up the speed of rumor propagation and increases the scale of propagation nodes.

Abstract:Deep reinforcement learning can enable the hexapod robot to obtain an optimal obstacle avoidance policy through training in a complex environment to complete the obstacle avoidance task.However,after the obstacles change in the environment,it often takes a lot of time to retrain the obstacle avoidance policy of the hexapod robot.Therefore,the dynamic obstacle avoidance of hexapod robot is still a challenging task.In this paper,a dynamic obstacle avoidance model of hexapod robot based on incremental reinforcement learning framework(IRL)is proposed to realize dynamic obstacle avoidance of hexapod robot.Firstly,the learned ob-stacle avoidance policy is relaxed through the policy relaxation mechanism to avoid policy falling into local optima.Secondly,the ef-fective states are given higher weights by importance weighting mechanism to update the obstacle avoidance policy to speed up the training process of the obstacle avoidance policy.The experimental results show that the proposed obstacle avoidance model of hexa-pod robot can adjust the obstacle avoidance policy of the hexapod robot more effectively than the baseline method after the obstacle changes.

Abstract:Human pose estimation typically involves detecting the joint positions and poses of the human body in an image.However,factors such as occlusion,complex textures,and changes in lighting make feature extraction difficult,making human pose estimation a challenging problem.In this paper,an improved ConvNeXt V2-based human pose estimation algorithm is pro-posed.A standard large-kernel convolution is decomposed into three convolutions,which are depthwise convolution(DW-Conv),depthwise dilation convolution(DW-D-Conv),and 1x1 channel convolution(1x1 Conv).Without changing the receptive field,this approach not only reduces computational complexity but also captures long-distance interdependent features.Additionally,by incorporating attention mechanisms into the network,this paper achieves adaptive spatial and channel dimensions,thereby improv-ing the network's feature extraction capabilities.Finally,experimental results demonstrate that the improved ConvNeXt V2 network outperforms the original ConvNeXt V2 network and other networks in human pose estimation to a certain degree.

Abstract:In recent years,with the increase of communication task requirements in broadband satellite systems,the resource scheduling problem such as time window conflicts has become more and more complicated,so this paper proposes an optimal re-source scheduling model for asynchronous time-division multiplexing(ATDM)forward link resources in broadband satellite commu-nication system based on multi-strategy improved ant colony algorithm(MIACO).Firstly,taking the system throughput as the opti-mization goal and the modulation coding mode,the number of multiple frames and the priority as constraints,a satellite resource scheduling model is established.Then,in order to solve the proposed model and consider the shortcomings of slow ACO conver-gence speed and easy to fall into local optimum,a MIACO based on tent mapping,ant colony factor adaptive update strategy and global pheromone update strategy is proposed.Finally,simulation experiments show that compared with GA and AGO,the proposed algorithm has strong optimization accuracy and convergence efficiency,and the scheduled business volume and system throughput of the proposed model are greatly improved.

Abstract:A new model DDH-YOLOv3 for the fast detection of occluded faces in complex scenes is proposed in this paper,under the complex conditions such as illumination changes,partial occlusion,and different face posture in natural scenes.Firstly,an improved multi-scale DRFBs visual field sensor module is introduced to enlarge the network receptive field and enhance the fea-ture extraction capability.Secondly,the CBAM module is specifically improved and ECA module is introduced to replace the chan-nel domain to reduce the complexity of the module,and parallel embedding is used to further improve the ability of the model to ex-tract occluded facial features.Finally,in order to diminish the missing rate of occlusive face,DIoU is used as the boundary frame loss function.The experimental results show that the improved model accuracy and FPS than YOLOv3 respectively increased by 5.69%and 28.81 s.At the same time,the detection performance is better compared with several other advanced target detection algo-rithms.Finally,the improved algorithms are ported to the ZYNQ platform for deployment and porting and showed good detection re-sults.

Abstract:This paper explores the influence of meteorological factors on the prediction of COVID-19.Considering the meteo-rological factors such as daily maximum temperature,daily minimum temperature,daily average temperature,daily average wind speed and the existing daily confirmed case data of COVID-19,a multivariate Long short-term memory neural network(LSTM)pre-diction model is constructed.At the same time,in order to improve the accuracy and operation efficiency of the prediction model,the dimension of the multivariate input data is reduced,and the meteorological factors with high correlation with the daily number of confirmed cases of COVID-19 are selected as the input of the model using Spearman correlation coefficient.The multivariable LSTM model with high correlation meteorological factors and existing daily confirmed case data of COVID-19 as model input has the best prediction effect.The multivariable LSTM prediction model established in this experiment can accurately predict the number of con-firmed cases of COVID-19,and has a strong generalization ability.

Abstract:Aiming at the problems of low recognition and detection ability of occluded objects,low accuracy and difficult rec-ognition of traffic signs of distant small targets in the detection of traffic signs by YOLOv5s model,a detection model based on im-proved YOLOv5s traffic signs is proposed.Firstly,by integrating the convolutional attention module(CBAM)and the weighted bi-directional feature pyramid network(BiFPN),the detection capability of YOLOv5s network model for traffic signs is strength-ened.Secondly,small target detection layer is added to splice shallow and deep feature maps.At the same time,image segmenta-tion and maximum suppression algorithm are used to improve the accuracy of small target traffic sign recognition.The experimental results show that compared with the original YOLOv5s model,the improved model is not only more accurate mAP@0.5,it has in-creased by 8.0%,6.5%and 2.7%respectively,and greatly improved the detection effect of small target traffic signs and traffic signs blocked by objects.The improved YOLOv5s model is superior to the original model and improves the accuracy of traffic sign detec-tion.

Abstract:Accurate molecular image recognition is essential for understanding the function of chemistry and its role in various biological processes,as well as for developing new drugs and treatments.In response to the low efficiency and accuracy of traditional methods for identifying molecular images,a combination of the EfficientNetV2 network and a multi-layer LSTM with a hybrid atten-tion mechanism is proposed to generate text from images and efficiently recognize molecular images.Multiple machine learning ex-periments are conducted with different encoders and decoders to obtain comparative results.The experimental results showes that compared to traditional machine learning models,this model has shorter training time and higher accuracy,achieving an accuracy of 93.1%in molecular image recognition.The combination of EfficientNetV2 and LSTM provides a promising method for molecular image recognition,simplifying the work of molecular image recognition and providing assistance for future chemical research.

Abstract:The base learner gradient boosting tree(GBDT)and regularized greedy forest(RGF)are used to replace the ran-dom forest and completely random forest in the cascade layer of the deep forest allows dynamic iterative optimization of global param-eters.Meanwhile,an adaptive structure is introduced in the cascade layer to update the input data of the next layer by weighting the classification results of the previous layer according to the correct rate,and an adaptive GBDT-RGF cascade.The adaptive GB-DT-RGF cascade deep forest model is established.The model can enhance the correct features and weaken the incorrect features in the previous layer during the cascade layer training to improve the classification accuracy of the model.In addition,for the stock ups and downs affected by multi-day trend,the stock factor data before the prediction date is restructured and weighted,and the weight-ed data is used to predict the stock ups and downs.Experiments show that the multi-factor quantitative stock selection model with adaptive GBDT-RGF cascading layer depth forest,which alleviates the problem of stock prediction lag,achieves an annualized re-turn of 36.55%in the CSI 300 stock pool from January 2020 to June 2022,with a cumulative return of up to 164%,exceeding the re-turns of models such as deep forest and random forest.

Abstract:The response characteristic drift of infrared detectors causes serious deterioration in imaging quality and system per-formance.Imaging simulation of infrared detector response characteristic drift can help to study methods to suppress the response characteristic drift and improve the performance of infrared imaging systems.However,existing research works are carried out from one aspect,and no unified response characteristic drift model of infrared detector has been formed,which cannot be effectively used in imaging simulation.Aiming at the problems above,an imaging simulation method of infrared detector response characteristic drift is proposed.A response characteristic drift model of infrared detector is established,which integrates the research results of infrared radiation,radiation transmission,photoelectric conversion and readout circuit in the imaging process of infrared detector.The imag-ing simulation processing steps of infrared detector response characteristic drift is designed.The simulation results show that this method can effectively simulate the response characteristic drift sequence image of infrared detector.

Abstract:To improve the classification effect of deep learning models on Tibetan text classification,a CBAt hybrid neural net-work-based Tibetan text classification model is proposed to classify 14 000 Tibetan texts in the dataset.In this study,the convolu-tional results of the traditional CNN model are directly used as the input of the BiLSTM model,and an Attention mechanism is intro-duced to increase the feature extraction of important information of Tibetan texts,so as to improve the classification accuracy.The ar-ticle also compares this dataset with traditional machine learning algorithms,single and hybrid neural network models,and the ex-perimental results show that the improved hybrid neural network model proposed in this study performs better in classifying Tibetan text.

Abstract:In order to solve the problem that the selection of traditional transformer state evaluation parameters is seriously af-fected by subjective selectivity and transformer index uncertainty,a transformer multi-source information fusion state evaluation method based on factor analysis and evidence theory is proposed.Among them,the transformer parameter index is extracted by prin-cipal component factor analysis method,and the state evaluation parameter system is determined by expert group experience and manual inspection parameters by variance maximization method.Then,the state evaluation of power transformer based on evidence theory is introduced to determine the membership function,and the membership function is modified.The data is iteratively opti-mized by feedback of fuzzy judgment results.Finally,the experiment of 240 MVA and 220 kV transformer verifies that the proposed method can objectively reflect the real state of transformer operation.

Abstract:In this paper,a new hybrid particle swarm algorithm is proposed to solve the problem that the traditional intelligent optimization algorithm has insufficient global search capability in the early stage of algorithm search and slow convergence speed in the later stage,which is easy to fall into local optimum.Firstly,the inertia weights of the particle swarm algorithm are dynamically adjusted using the cosine adaptive strategy,which effectively improves the search performance in the early and late iterations of the algorithm.Secondly,the explosion mechanism of the fireworks algorithm is introduced to perform explosion search at the historical optimal position of each particle to improve its local fine search capability.Finally,the health degree of each particle is detected,and the lazy particles below the health degree threshold are eliminated from the population by using the idea of"survival of the fit-test"of the wolf pack algorithm,and the same number of particles are randomly generated into the population by introducing the wolf pack update mechanism,which improves the population diversity and makes the algorithm have a strong pioneering ability.The proposed hybrid particle swarm algorithm is proved to be feasible,reasonable and efficient in solving the FJSP problem by conduct-ing simulation experiments on the FJSP arithmetic case and comparing with other algorithms.

Abstract:The feature point algorithm plays an important role in image processing.The ORB(Oriented FAST and Rotated BRIEF)feature point extraction algorithm is widely used because of its rotation invariance and fast computation speed.The tradition-al ORB feature point algorithm can not detect feature points easily in light-transformed scenes,and the running time is not enough for a strong real-time system.Based on the principle of ORB feature point algorithm,this paper improves the original ORB algo-rithm by using methods such as reducing the area of detection area,optimizing the detection process,optimizing trigonometric func-tion solution,dynamic gray threshold,etc.It improves the speed of ORB feature point detection algorithm and improves the applica-bility of ORB feature point detection algorithm under different lighting conditions.Finally,experiments show that the improved ORB algorithm takes less time on average than the algorithm in OpenCV library,and can extract more feature points in scenes with vary-ing brightness.The improved algorithm is more suitable for real-time systems such as visual SLAM(Simultaneous Localization and Mapping)and systems that use ambient light to easily transform.

Abstract:In this paper,an improved whale optimization algorithm is proposed to solve the maximum completion time prob-lem of flexible job shop scheduling.ROV value mapping is used to map machine selection and job scheduling to the continuous opti-mization algorithm,and the Tent chaos strategy is used to initialize the population at the initialization stage,based on WOA,the nonlinear control theory parameter balanced global search and local search are introduced,the Lévy flight strategy is also introduced to expand the population exploration space,and the TS is embedded in the update phase of WOA to enhance the local search capa-bility.Finally,the improved whale optimization algorithm is proven to be able to find a better solution by testing a standard example of a flexible workshop.

Abstract:In order to solve the problem of huge workload brought to doctors by huge lung CT data,this paper proposes a lung nodule detection method based on improved U-Net network.The method in this paper takes the U-Net network as the basic frame-work.Firstly,residual modules are added to the coding part to increase the depth of the network.Then,the deep supervision module is added to the decoding part,so that the shallow layer can be more fully trained.Finally,the attention mechanism module is used between the encoding and decoding parts to make the model more focused on learning useful information.The experimental results show that the method in this paper has better performance than other detection methods,and the Dice coefficient,sensitivity and specificity on the public LUNA 16 dataset are 0.8109,90.62%and 98.97%,respectively.

Abstract:To address the issue of reduced accuracy and stability caused by energy dissipation in micro-electromechanical sys-tems(MEMS),this paper focuses on the dynamic characteristics and energy dissipation mechanisms of thermoelastic strongly cou-pled structures.It investigates topology optimization techniques to achieve minimal thermoelastic damping.Based on these optimiza-tion techniques and the Matlab development environment,a design software system for optimizing MEMS thermoelastic coupling structures is developed.This system integrates theoretical models,optimization algorithms,and solution techniques,and its compu-tational accuracy is verified through numerical examples.The software provides tool support for the design of high-quality and highly stable MEMS thermoelastic strongly coupled structures.