Currently,Wi-Fi-based wireless personnel perception technology is widely used in anti-intrusion security monitoring,human health care,gait recognition and other fields,regarding this,this paper proposed a non-cooperative human behavior recognition method.The channel state information (CSI) of Wi-Fi signals can be used to recognize five dynamic activities:walking,sitting-standing up,squatting,jumping and falling.The method uses a SIMO system to collect CSI data,and after performing pre-processing on the CSI amplitude and phase respectively,implements a three-step computational cost reduction mechanism:subcarrier fusion,rejection of bad data link based on mobile variance threshold,and data segmentation of dynamic time window based on wavelet transform.Then activity features are extracted and extended from the time domain to the frequency domain.By analyzing the characteristics of the Doppler power spectrum,the utilization of the CSI signal is improved.Experiment results show that the overall recognition rate increases with the use of feature dimensions.Optimized by two rounds of voting,the combined classifier weighted voting method is increasing the overall recognition rate of five dynamic activities to 90.3%.And compared to RSSI,the advantages of CSI in the field of human behavior recognition are more prominent.

Traffic light detection and recognition technology can help drivers make correct driving decisions,reduce traffic accidents,and provide security for unmanned driving.Aiming at the technical difficulties such as the complex and variable traffic light detection scene,and targets typically account for a very small percentage of the dataset images,a fast detection and recognition algorithm for traffic light based on deep learning was proposed.The overall framework consists of three parts:heuristic-based image pre-segmentation,which is used to narrow the search range and improve the relative size and detection accuracy of the traffic light panel in the input images;detection and recognition based on deep learning,using convolutional neural networks to detect and identify traffic lights accurately;NMS (Non-Maximum Suppression) algorithm,which is used to remove the repeated detections of the previous stage.The proposed Split-CS-Yolo model achieves 96.08% mAP and 2.87% miss detection rate on the LISA dataset.Compared with other methods of the Yolo series,it not only has higher accuracy and lower missed detection rate,but also reduces the model size to 8.6% of the original Yolov2,thus increasing the detection speed by 63%.

The essence of dynamic routing in capsule network is the implementation of clustering algorithm.Considering that the clustering method used in the previous capsule network requires the data to meet certain distributions to achieve the best effect while features of image are complicated,a more universal fuzzy clustering algorithm was taken as the feature integration scheme to replace the old in this paper.And an activation value using information entropy to measure the indeterminacy was added to the model,so as to distinguish the significance of capsule features at the same layer.Meanwhile,drawing on the idea of feature pyramid network,the features of different capsule layers are sampled to the same size to fuse and then are trained independently.Experimental results based on the Keras framework show that the capsule network with new structure has higher recognition accuracy on MNIST and CIFAR-10 than the original capsule network.The contrast experiments prove great potential of fuzzy clustering algorithm applying on capsule network,which alleviates the limitation of the clustering algorithm in the original capsule network.The results also prove that the features of different layers in the capsule network can be fused to be more informative and expressive.

Zero-shot classification algorithm does not need to label the sample of unseen classes to be recognized,so it can greatly reduce the cost of practical application,which has attracted wide attention in recent years.The problem of structure difference between word vectors and image feature prototypesseriously affects the zero-shot classification performance of remote sensing scenes.Based on the complementarity among different kinds of word vectors,the remote sensing scenes zero-shot classification algorithm based on word vectors fusion,named coupled analysis dictionary lear-ning method,was proposed.Firstly,the sparse coefficients of different kinds of word vectors are obtained by the more efficient analysis dictionary learning to reduce the redundant information.Then,the sparse coefficients are concatenated and denoted as the fused word vectors,and a structure alignment operation is performed based on the image feature prototypes to reduce structural differences by embedding the fused word vectors into image feature space.Finally,the image feature prototypes of the scene classes unseen are calculated,and the nearest neighbor classifier is employed to complete the classification in the image feature space.Quantitative experiments of the fusion of multiple semantic word vectors were carried out on UCM and AID datasets.At the same time,two real remote sensing images were qualitatively tested with RSSCN7 datasets as the seen dataset.Auantitative experiments obtaines the highest overall classification accuracies of 48.40% and 60.23% on UCM and AID,which respectively exceeds the typical comparative methods by 4.80% and 6.98%.In qualitative experiments on two real remote sensing images ,the algorithm also obtaines the best zero-shot classification performance.The experimental results show that the fused word vectors are more consistent with the prototypes in image feature space,and the zero-shot classification accuracies of remote sensing scenes can be significantly improved.

At present,cardiovascular diseases have become the main cause of global non-communicable death,death toll accounts for about one third of the total toll of death in the world,and the number of patients is increasing year by year.Wearable devices is used to automaticaly classify electrocardiogram to facilitate the early monitoring and prevention of cardiovascular diseases for patients.With the rise of edge machine lear-ning and federated learning ,small machine learning models have become a hot issue.According to the characteristics of wearable electrocardiogram equipment such as low configuration,low power consumption and personalization,this paper studied a lightweight network model based on LSTM,and used adaptive algorithm to optimize the ECG classification model of individual patients.The experiment is conducted by using the MIT-BIH open dataset.And compared with the current studies on the detection performance of VEB and SVEB,the experiment results show that the proposed algorithm has simple model structure and high classification performance,which can meet the requirement of ECG monitoring for patients by wearable devices.

Aiming at the problem of efficient retrieval of massive computer aided design(CAD) models in enterprise product manufacturing process,this paper studied a retrieval algorithm based on the content feature of two-dimensionalCAD models,and constructed a retrieval system prototype which can be used for DXF format CAD source file model base.Firstly,through the analysis of the DXF file structure of the two-dimensional CAD model,the rule of the primitive in the model is studied and the shape reconstruction is carried out.Secondly,according to the features of primitive,three kinds of content feature extraction methods are proposed,which are based on statistical histogram,two-dimensional shape distribution and Fourier transform.Finally,a multi-feature fusion framework based on unsupervised learning and similarity calculation method are designed to extract the fusion feature descriptor of the model and realize the retrieval of two-dimensional CAD model.Experiments show that the fusion features extracted in this paper contain more abundant content features and are more effective than single features.The system can be directly used in product customization,product design reuse and other aspects to help enterprises further improve the ability of intelligent manufacturing.

With the development of ubiquitous computing,more and more sensors are installed in our daily applications.As a result,the demand for time series data processing is very high.The similar pattern which appears in time series data several times are called time series motif.Motif contains huge amounts of information in time series data.Motif discovery is one of the most important work in motif analysis.State-of-art motif discovery algorithm cannot find proper motif based on domain knowledge.As a result,such algorithm cannot find most valuable motif.Aiming at this problem,this paper used domain distance to evaluate the similarities of subsequences based on domain knowledge.By using the new distance,this paper developed a branching method to discovery motif with variable length.Several data from real life are used to test the performance of the algorithm.The results show that the proposed algorithm can find motif with domain knowledge accurately.

In order to adaptively decompose the image and accurately describe the distribution state of the decomposition coefficients,a new image denoising algorithm based on fast and adaptive bidimensional empirical mode decomposition algorithm was proposed.Firstly,the algorithm performs fast and adaptive bidimensional empirical mode decomposition on the image.By determining the number of noise-dominated subband after decomposition,the noise-dominated subband coefficient distribution is further modeled by the normal inverse Gaussian model.Then the Bayesian maximum posteriori probability estimation theory is used to derive the corresponding threshold from the model.Finally,the optimal linear interpolation threshold function algorithm is used to complete the denoising.The simulation results show that for adding Gaussian white noise images of different standard deviation,the average signal-to-noise ratio is improved by 4.36dB,0.85dB,0.78dB and 0.48dB,respectively,compared with sym4 wavelet denoising,bivariate threshold denoising,pro-ximity algorithms for total variation,and overlapping group sparse total variation algorithm.Structural similarity index is also improved with different degrees,which shows it can effectively preserve more image details.The experimental results show that the proposed algorithm is superior to the comparison algorithms in terms of visual performance and evaluation index.

The imaging factors such as illumination,view,obstruction and noise would bring great changes to pedes-trian’s appearance under the cross-view condition in person re-identification,then it becomes very difficult to identify the target from candidates.Using the re-ranking algorithm can optimize the re-identification’s result,but it can make the task time-consuming and expensive.What’s more,it is easy to introduce the noise during the process of re-ranking,which in turn affects the accuracy of re-identification.To solve the problem,this paper presented a re-ranking method based on bidirectional KNN for person re-identification.First,it utilized the pre-training and fine-tuning strategy to extract the deep features of pedestrian.Then,it choosed an appropriate metric function (XQDA,KISSME) to measure the distance of features.Finally,accor-ding to the bidirectional KNN relation between the query and candidates,the Jaccard distance was calculated and aggregated with the original distance to guide the re-ranking.Experiments on the datasets of CUHK03,Market1501 and PRW show that the re-ranking algorithm proposed in this paper can improve the accuracy of re-identification on the basis of the original method,and the improvements are 12.2% and 13.4% in the two evaluation indexes of Rank1 and mAP respectively.The experimental data indicates that the re-identification algorithm based on bidirectional KNN can effectively reduce the probability of noise during the re-ranking,and then improve the accuracy of re-identification.

Object detection is an important research direction in the field of computer vision.In recent years,deep lear-ning has achieved great breakthroughs in object detection which is based on the video.Deep learning has powerful ability of feature learning and feature representation.The ability enables it to automatically learn,extract and utilize relevant features.However,complex network structure makes the deep learning model have a large scale of parameter.The deep neural network is both computationally intensive and memory intensive.Single Shot MultiBox Detector300 (SSD300),a single-shot detector,produces markedly superior detection accuracy and speed by using a single deep neural network.But it is difficult to deploy it on object detection systems with limited hardware resources.To address this limitation,the fusing method of weight pruning and filter pruning was proposed to reduce the storage requirement and inference time required by neural networks without affecting its accuracy.Firstly,in order to reduce the number of excessive weight parameters in the model of deep neural network,the weight pruning method is proposed.Network connections is pruned,in which weight is unimportant.Then,to reduce the large computation in convolution layer,the redundant filters are pruned according to the percentage of effective weights in each layer.Finally,the pruned neural network is trained to restore its detection accuracy.To verify the effectiveness of the method,the SSD300 was validated on caffe which is the convolutional neural network framework.After compression and acceleration,the storage of SSD300 neural network required is 12.5MB and the detection speed is 50FPS.The fusion of weight and filter pruning achieves the result by 2× speed-up,which reduces the storage required by SSD300 by 8.4×,as little increase of error as possible.The fusing method of weight and filter pruning makes it possible for SSD300 to be embedded in intelligent systems to detect and track objects.

The distribution of urban water bodies is of great significance for people to understand the geographical phenomena such as the urban water circulation and the Heat-island Effect.It is common to obtain information by using high-resolution images for water extraction and water mapping.However,automatically extraction of water bodies by using the high-resolution images still is difficult for the complex background of the urban area,fewer spectral channels provided by the high-resolution images and the uneven distribution of water bodies in the images.This paper proposed an automatic extraction method of urban water bodies in complex background based on high-resolution images.First,adaptive threshold is selected for segmentation to gain the initial region of water,since water has a low gray value of the near infrared channel.Next,on the initial region,a buffering algorithm are used to obtain the target region of water extraction,and gauss mixture model and an expectation maximization algorithm is used to improve the distributionpara-meters of water.Then,the water bodies are extracted automatically using the maximum likelihood method with these parameters.As for the large number of shadow elements mixed in the rough extraction,a fusion features method is proposed to eliminate those noise points and obtain more accurate extraction result.The experiment results of water extraction in Jinshan show that the proposed method can effectively extract the structure of water bodies with small proportion in the experimental images,and perform well with high accuracy comparing to the commonly used automatic extraction algorithms.

Emotion analysis is a fundamental task of natural language processing(NLP),and the research on single modality (text modality) has been rather mature.However,for multi-modal contents such as videos which consist of three modalities including text,visual and acoustic modalities,additional modal information makes emotion analysis more challenging.In order to improve the performance of emotion recognition on multi-modal emotion datasets,this paper proposed a neural network approach based on multi-task learning.This approach simultaneously considers both intra-modality and inter-modality dynamics among three modalities.Specifically,three kinds of modality information are first preprocessed to extract the corresponding features.Secondly,private bidirectional LSTMs are constructed for each modality to acquire the intra-modality dynamics.Then,shared bidirectional LSTMs are built for modeling inter-modality dynamics,including bi-modal (text-visual,text-acoustic and visual-acoustic) and tri-modal interactions.Finally,the intra-modality dynamics and inter-modality dynamics obtained in the network are fused to get the final emotion recognition results through fully-connected layers and the Sigmoid layer.In the experiment of uni-modal emotion recognition,the proposed approach outperforms the state-of-the-art by 6.25%,0.75% and 2.38% in terms of text,visual and acoustic on average respectively.In addition,this approach can achieve average 65.67% in accuracy in multi-modal emotion recognition tasks,showing significant improvement compared with other baselines.

Fire detection based on image processing technology is an important branch in the field of fire monitoring in recent years.Aiming at the fire detection of open environment,using the dynamic and static characteristics of smoke and flame generated during the fire,the two-stream convolutional neural network is used as the theoretical basis to detect the fire.The two-stream convolutional neural network uses spatial and temporal streams to extract spatial information and temporal information in the video respectively.However,the information in the early stage of the fire is weak and the features are not obvious enough.In order to improve the initial recognition rate,a spatial enhancement network was proposed as the spatial stream of the two-stream convolutional neural network to extract and enhance the spatial information of the video.The spatial enhancement network simultaneously convolves the current frame V_t and the previous frame V_t－1,subtracting the convolution features of the V_t image with the convolution features of the V_t－1 image,to preserve the difference of the convolution features,and adding the convolution features difference to the convolution features of the current frame V_t,thereby enhance the spatial features convolution of the current frame V_t－1.Temporal stream of two-stream convolutional network convolves the optical flow image V_t′ of the current frame to get the temporal features.Finally,the enhanced spatial and temporal features are fused to classify.The experimental results show that the improved two-stream convolutional network has a 6.2% higher recognition rate than the original two-stream convolutional network,and achieved 92.15% recognition rate on the public dataset,indicating the effectiveness and superiority of the proposed method.Comparing with other methods,the network structure is designed lower but achieves good results,improves the identification accuracy of fire and smoke as well as realizes the early warning of fire,shorten detection time.

In view of the disadvantages of the traditional methods,such as low accuracy,slow classification speed,and a great difference in the recognition accuracy of different types of track defects,a new method of track defect image classification based on improved ant colony algorithm was proposed.The track defect image is preprocessed,the vertical projection method is used to extract the track surface area,the fuzzy theory and the hyper entropy theory are combined to obtain the best segmentation threshold,and the image segmentation is completed.Combined with adaptive threshold Canny edge detection operator and Hough transformation method,the rail defect part is determined.The edge details of defects are improved to make the contour of track defects more obvious.On the basis of this,the basic ant colony algorithm is analyzed,the characteristic similarity is used as a discriminant function,and the improved ant colony algorithm is used to classify the track defect image.Experimental results show that the classification accuracy and classification speed of the proposed method are high.

Face hallucination reconstruction refers to the process of reconstructing high-resolution enhanced face from a low-resolution image.Most of the traditional methods assume that the input image is aligned and noise-free.However,the super resolution performance will decrease when the input facial image is unaligned and affectedby noise.This paper proposed an effective single image super resolution method for unaligned face images,in which the learning-based hierarchical clustering regression approach is used to get better reconstruction model.The proposed face hallucination methodcan be divided into clustering and regression.In the clustering part,a dictionary is trained on the whole face image with tiny size,and the training images are clustered based on the Euclidean distance.Thus,the facial structural prior is fully utilized and the accurate clustering result can be obtained.In the regression part,to reduce the time complexity effectively,only one global dictionary needs to be trained during the entire training phase whose atoms are taken as the anchors.In particular,the learned anchors are shared with all the clusters.For each cluster,the Euclidean distance is used to search the nearest neighbors for each anchor to form the subspace.Moreover,in every subspace,a regression model is learned to map the relationship between low-resolution features and high-resolution samples.The core idea of this method is to utilize the same anchors but different samples for clusters to learn the local mapping more accurately,which can reduce training time and improve reconstruction quality.The results of comparative experiments with other algorithms show that the PSNR can be increased by at least 0.39 dB and the SSIM can be increased by 0.01 to 0.18

Existing corresponding algorithms will cause error mappings since geometric signatures can’t remain stable and highly similar under different poses.This paper focused on corresponding 3D dynamic surface based on space-time constraints.Firstly,this algorithm constructs the energy optimization function according to the non-rigid deformation model.Secondly,sparse correspondence is computed by optimizing the energy function.Finally,this algorithm of surface sampling and isometric mapping is used to solve the dense matching problem.In experimental part,the analysis and quantification of different 3D motions are carried out,and it turns out that this algorithm can improve the correspondence accuracy.

Fractal image encoding with the high compression ratio can maintain a good quality of reconstructed image.However,there are some disadvantages such as high computational complexity and long encoding time.Therefore,based on the definition of a new sub-block feature called sum of frame and point,combined with the smoothing characteristics of continuous wavelet transform,an image compression encoding on the basis of wavelet transform and fractal was proposed.This algorithm makes full use of the correlation of sub-bands,so as to improve the quality of reconstructed ima-ge.And it converts the global search into the nearest neighbor search to shorten search range and reduce encoding and decoding time.The simulation results show that compared with the basic fractal algorithm and other algorithms,the new algorithm has better performance.In addition,it not only shortens the encoding and decoding time,but also improves the reconstructed image quality

In order to solve the problem of poor binaryzation quality caused by uneven illumination and uncontrollable environment in OCR applications of mobile terminals,this paper proposed an adaptive multi-level threshold binaryzation method based on integral graph.First,a specific sliding window is set by focusing on the points to be calculated.The normal threshold is the mean value of the sliding window where the current point is located.The two front sliding windows are weighted according to the Gauss function,and then the relaxation factor is obtained according to the weights.The relaxation threshold of pixels are obtained based on the evaluation of the relaxation factor and illumination condition.Experiments were carried out in typical mobile environments such as irregular shadows,multi-level illumination and linear light changes.Lenovo ZUK Z2 Pro is used as the test equipment.The average recall of the algorithm is 95.5% and the average accuracy is 91%.The recognition accuracy of this algorithm is 96.8%,98.2% and 93.2% respectively in the environment of irregular shadow,multilevel illumination and linear light change.The result shows that the proposed algorithm has strong robustness and adaptability,and can meet the requirement of image preprocessing in the OCR application of mobile terminal

The evaluation method of renal function in clinical renal dynamic imaging depends too much on manual acquisition of ROI (Region of Interest)and has low time efficiency.In order to solve this problem,this paper proposed anautomatic quantitative assessment method for medical renal dynamic imaging.Firstly,the images of renal dynamic imaging at different stages are pretreated.Secondly,an improved level set model is utilized to obtain the ROI of the renal function imaging.The ROI is obtained by morphological methods,then the ROI of the aorta in the renal perfusion imaging is located and obtained.Finally,GFR(Glomerular Filtration Rate) is calculated according to the Gates method,and the time-radioactivity curve is plotted based on the radioactivity counts in ROI,so as to achieve integrated and automated assessment for renal function.The results of clinical trials show that the proposed automatic assessment method can improve the automation level in a short period of time and raise the assessment accuracy,which provide effective help for clinical diagnosis and adjuvant treatment

Due to the influence of strong light,the images acquired at night have high contrast,the same situation also appears in backlit images collected in the daytime.Contrast enhancement method is usually applied to the images for obtaining images with favorable contrast.Whereas,over-enhancement commonly occurs in bright regions.Accordingly,in order to solve the problem of over-enhancement for high contrast images,a Retinex based low light image enhancement algorithm through adaptive reflection component estimation and logarithmic image processing subtraction post-proces-sing was proposed.The algorithm mainly includes into two parts:reflection component estimation and logarithmic image processing subtraction (LIPS) enhancement.First,adaptive parameter bilateral filters are used to get more accu-rate illumination layer data,instead of Gaussian filter.Moreover,the weighting estimation method is used to calculate the adaptive parameter to adjust the removal of the illumination and obtain the reflectance by just-noticeable-distortion (JND)factor.In this way,it can effectively prevent the over-enhancement in high-brightness regions.Then,the LIPS method based on maximum standard deviation of the histogram is applied to enhance reflectance component part,where the interval of the parameter is according to the cumulative distribution function (CDF).Experimental results demonstrate that the proposed method outperforms other competitive methods in terms of subjective and objective assessment

Aiming at the problem of low target detection accuracy and poor system robustness in ship image target detection,an improved YOLO v2 algorithm was proposed to detect ship image targets.The traditional YOLO v2 algorithm is improved by clustering the target frame dimension,optimizing the network structure,multi-scale transformation of input image,so as to better adapt to the ship target detection task.The test results show that the mean Average Precision (mAP)of the algorithm is 79.1% when the input image size is 416×416,and the detection speed is 64 frames per se-cond (FPS),which can satisfy the real-time detection and exhibit high precision and strong robustness for small target detection

In view that the existing DCA (dendritic cell algorithm) relies heavily on domain knowledge and artificial experience defining antigen signals in fault detection application,and a single antigen anomaly evaluation method can’t reflect the overall health condition of system,this paper proposed a fault detection method based on immune homeostasis mechanism－IHDC-FD.First of all,in order to solve problem that the danger signal definition is not explicit in actual application,by introducing body’s immune homeostasis mechanism,the change that breaks the homeostasis is consi-dered to be the danger source of system.Therefore,the method of antigen signal of DC adaptive extraction from the change of system state by numerical differential method is proposed.Secondly,the concentration of specific cells within the tissue is the critical factor that can reflect the health of body,and in order to keep healthy,the body’s immune homeostasis has to be maintained.So,by reference to the activation and suppression mechanism of body’s immune homeostasis,the Th and Ts cell concentration which maintain the immune homeostasis is regarded as the evaluation indicators of system imbalance,and once the system lose balance,a fault occurs.Finally,the performance of our method is tested by using step,random and slow drift faults on TE benchmark.Compared with the original DCA,the results show that IHDC-FD not only improves the adaptability of DCA,but also increases the average of fault detection rate by 9.93%,decreases false alarm rate by 230.4% and decreases delay time by 101.2% on the three types of faults testing.Therefore,the IHDC-FD method based on immune homeostasis mechanism has a large improvement than the original DCA on detection performance and adaptability,and it is effective and generality

As an important information carrier,image is indispensable in life,and how toretain and acquire the information in the image to the greatest extent has been a big topic for a long time.In recent years,radial basis function (RBF) interpolation has become a new effective method to solve the problem of scattered data interpolation.In the image magnification based on radial basis function,the values of different parameters have a great influence on the magnified ima-ge.The appropriate fidelity index is particularly critical for the image quality evaluation and the study on the parameters.This paper mainly presented the definition of fidelity index for image magnification based on the multilevel feature of image and the radial basis function of the block matrix,which consists of the global distortion index and the edge distortion index.The experimental results show that the definition of fidelity index is effective.Furthermore,the correlations between the parameters of MQ,inverse MQ and the Gauss radial basis functions and the image texture amplification mechanism were studied.

Since traditional multi-focus image fusion algorithm has the problem of low contrast ratio,this paper presented a multi-focus image fusion algorithm based on improved weighted method and adaptive pulse coupled neural network (PCNN) in Shearlet domain.Firstly,the source images are decomposed by Shearlet transform to generate a low-frequency subband and a series of high-frequency subbands with different scales in different directions,then the weighted sum of the low-frequency subbands and the absolute value of the difference of the low-frequency subbands are conducted,the weight is calculated by the average gradient,and finally the fused low-frequency subbands are obtained.At the same time,the high-frequency subbands are fused by adaptive PCNN fusion rule,the motivation for PCNN is calculated by sum-modified Laplacian,the linking strength for PCNN is adaptively calculated by the regional spatial frequency of each source images,and the fused high-frequency subbands are obtained according to the ignition map of PCNN.Finally,the fusion image is acquired by the Shearlet inverse transform.One group of artificial simulated multi-focus images named Cameraman and three groups of real multi-focus images named Pepsi,Clock and Peppers are selected respectively for experiments,seven different fusion methods are chosen as a comparison,and four common quality evaluation indexes are used to evaluate the fusion images objectively.The experimental results show that the proposed method has good performance both on subjective vision and objective evaluation.

For the overall quality of the sentence describing the generated image is not high in the current image description task,and an image description model fusing word2vec and attention mechanism was proposed.In the encoding stage,the word2vec model is used to describe the text vectorization operations to enhance the relationship among words.The VGGNet19 network is utilized to extract image features,and the attention mechanism is integrated in the image features,so that the corresponding image features can be highlighted when the words are generated at each time node.In the decoding stage,the GRU network is used as a language generation model for image description tasks to improve the efficiency of model training and the quality of generated sentences.Experimental results onFlickr8k and Flickr30k data sets show that under the same training environment,the GRU model saves 1/3 training time compared to the LSTM model.In the BLEU and METEOR evaluation standards,the performance of the proposed model in this paper is significantly improved.

Due to the individual difference of surface electromyography (sEMG),an individual person always needs long-time pre-training for obtaining his own accurate classification model when using sEMG as control source of external equipment.For solving this problem,on the basis of the original KKT-SVM incremental learning method,a new SVM incremental learning algorithm (D-ISVM) based on DBSCAN density clustering was proposed and it was applied in the on-line sEMG hand gesture recognition.Firstly,considering that the new samples and initial non-SV samples can affect new SV set,the closeness of sample distribution is analyzed and clustered according to DBSCAN,and the new samples and initial non-SV samples which are close to initial SV set are selected.Then,these samples are furtherselec-ted based on core point and distance between samples and hyperplane.Finally,all selected samples and initial SV set are trained together to obtain new SV set.The experimental results show that,compared with general algorithms,the proposed D-ISVM incremental learning algorithm can achieve higher classification accuracy and further improve the learning speed of classification model.This method can effectively solve the individual difference problem during the on-line sEMG hand gesture recognition.

In order to solve the problem that the speed and accuracy need to be taken into account simultaneously when conducting multispectral image matching,this paper improved the SIFT algorithm from the following several aspects.Aiming at the problems such as the slow matching speed and low matching rate caused by high dimension of feature descriptors,this paper improved the structure of feature descriptors to reduce the dimensions of descriptors.In the aspect of SIFT feature matching,firstly,the feature point is determined as the maximum point or minimum point according to the trace of Hessian matrix,which can narrow subsequent search range for the feature vector matching.Then,the partial matching point pairs are eliminated based on the position information of feature points.The experimental results show that the improved algorithm not only preserves the invariance advantages of the traditional algorithm,such as rotation and brightness,but also can effectively reduce the running time,and improve the matching rate on a certain extent.

Morphological methods use structural units to measure and extract the target shape in the image to achieve the purpose of image analysis and processing,and these methods have been widely used in binary and grayscale image processing.In order to extend the gray morphology to the color image,this paper defined the color similarity in the HSI color space and proposed the color morphological image processing method.Firstly,the color similarity measure is defined by combining hue,saturation and intensity in the HSI space to characterize the similarity degree between color vectors.Then,the new type of color morphology is constructed by using color similarity,including the basic operations such as dilation,erosion,opening and closing.Finally,the morphological basic operations combined with color similarity are applied to extract color image edges.Experiments make in-depth analysis and research on the color morphological image processing performance,and it can be found that the color morphology operation is relatively better when the parameter k≤0.05.The experimental results show that the proposed method has the ability of smoothing the edge of the color target and the edge extraction performance.At the same time,it also shows the practicability and effectiveness of the image processing.

Quaternion wavelet transform (QWT) is a new multi-scale transform tool which can provide both amplitude and phase information.In this paper,copula model is used to capture the correlation of QWT coefficients,and a novel image fusion method based on QWT and Copula modelwas proposed.First,QWT is performed on the source images.Second,the dependency among the magnitude-phase of high frequency subbands and the corresponding phase of low frequency phase is established by Copula models.Next,a choose-max fusion rule based on the comprehensive feature constructed by the regional energy of Copula joint probability density,the gradient of phases,the QWT coefficient energy and the local contrast,is proposed for high frequency subbands.A choose-max fusion rule based on the comprehensive feature constructed by gradient and local variance of low frequency phases is proposed for low frequency subbands.Finally,the fusion image is obtained by inverse QWT.Experimental results demonstrate that the performance of the proposed method is superior to the traditional fusion methods.

Gram matrix,a method to get the inner product in simple terms,was commonly used for image style extraction in the style-transfer techniques.The Gram matrix can only extract the static features,but it is completely unconstrained to the spatial sequence of objects in the picture.This paper proposed a fast neural style transfer method with space constraints.First,the residuals are used to redesign the transform network of fast neural style transfer.Then,the method of spatial offset is used to transform the Feature map.Feature map T(a^l) are used for Gram matrix computation to get the cross-correlation,which contains the spatial information.That is to say,it can constrain the object’s spatial sequence in the picture.Finally,experiments show that the method’s ability of space constraint is better than traditional method,and the stylized image with better effect can be quickly obtained.