In modern medicine,the image of the brachial plexus segmentation and recognition is optimized by contrast enhancement to help the physician identify the disease and tumor.Brachial plexus block is a commonly used method of local anesthesia in upper limb surgery and postoperative care.In order to accurately determine the position of thebrachialplexus,the hospital extensively applies ultrasound equipment to detect and locate the nervous system.This paper described the accurate recognition and segmentation of brachial plexus in ultrasound dynamic images based on deeplear-ning and neural network,and optimized the display of ultrasound images through adaptive contrast enhancement in the cut-out images.The experiment data come from the Beijing Jishuitan Hospital,which are divided into ultrasound images of patients and corresponding pictures of benign malignancies.The enhanced contrast algorithm was used to process the extracted features.The experimental results show that this algorithm enhances the contrast of the image and the accuracy of the displayed content.

The key problem of stereo vision is to obtain accurate disparity values through stereo matching algorithms.However,most existing stereo matching algorithms are unable to obtain accurate and correct disparities in low-texture regions.In this paper,in order to solve the problems of low matching accuracy of low texture regions and large computational complexity of high-precision semi-global matching algorithm,a stereo matching algorithm based on adaptive sliding window was proposed.The cost volume is calculated by AD-Census transform firstly.The shape of the aggregate window and the weight of the pixels are adjusted for different regions.The cross-scale cost aggregation framework conforming to the human visual feature is used to obtain the aggregate cost volume.Finally,the “winner take all strategy” is used to obtain the final disparity maps.Experiments show that the mismatch rate of the algorithm in low-texture regions decrease form 5.8% to 21.68%,which is lower than that of the traditional scheme,and the computation time is shorter than the semi-global algorithm.

Faces in the wild are huge in number and more close to life,and the recognition of facial attributes is a valuable research.A face attributes recognition method named RMLARNet (Regional Multiple Layer Attributes Related Net) was proposed for faces in the wild,which explores a new feature extraction method and attributes relationship.The processing steps of this method are as follows:1)Feature extraction is based on the regional parts of image.2)Features are extracted from different layer of Inception V3,and they are concatenated to get the final face feature.3)An attributes relationship related network is used for attributes recognition.The experiment is conducted on a balanced CelebA- data set which is a subset of CelebA,and this method outperforms state-of-the art methods.

The large-scale smart driving scene reconstruction can feedback the surrounding road traffic environment information for the vehicle control system in the vehicle driving environment,and realize the visualization of the environmental information.At present,the existing three-dimensional reconstruction scheme is mainly oriented to thestructuredscene,and it is difficult to meet the real-time performance required by the smart driving system while ensuring a certain precision which can make when the three-dimensional reconstruction of the large-scale unstructured smart driving scene is performed.In order to solve this problem,a three-dimensional scene reconstruction method based on binocular vision is proposed.Firstly,by optimizing the stereo matching strategy,the stereo matching efficiency is improved,and then the uniform distance feature point extraction algorithm RSD is proposed to reduce the time consumption of 3D point cloud computing and triangulation,and the real-time performance of large-scale smart driving scene reconstruction is improved.The experimental results prove the effectiveness of this algorithm,which can be used to reconstruct the scene of large-scale smart driving scene,and can meet the demand of intelligent driving system in real-time.

The thesis used the geometric characteristics of collinear points to get the intrinsic parameters of the came-ras.Firstly,the homographic matrix between space collinear points and its image points is used to get the linear constraints of the intrinsic parameters and the intrinsic parameters for multiple cameras.Then,according to the coordinates of collinear points before and after movement in each camera,the rotation matrix and translation vector of the camera relative to the reference camera are obtained,and the outside parameters of the cameras are solved.Finally,simulation data and real image experiments show the feasibility and effectiveness of this method.

Face clustering which aims to automatically divide face images of the same identity into the same cluster,can be applied in a wide range of applications such as face annotation,image management,etc.The traditional face clustering algorithms can achieve good precision,but low recall.To handle this issue,this paper proposed a novel clustering algorithm with triangular constraints and context constraints.The proposed algorithm based on conditional random field model takes triangular constraints as well as common context constraints into accountin images.During the clustering iteration and after preliminary clustering,maximum similarity and people co-occurrence constraints are considered to merge the initial clusters.Experimental results reveal that the proposed face clustering algorithm can group faces efficiently,and improve recall with the high precision,and accordingly enhance the overall clustering performance.

To address the problem of dynamic sequential recovery for Chinese handwritten,a handwritten drawing order recovery model based on deep learning method was designed.First,the handwritten image is preprocessed by coordinate regularization,refinement,and interruption of intersections,then the preprocessed image and the corresponding written coordinate sequence are used to generate the sample of the network.The sample consists of a static handwritten image and a heat map label containing the font writing order.The model uses an end-to-end convolutional neural work.Finally,the trained network model is used to predict the static handwritten image to get the original writing order of the font.The experimental results show that the method can effectively recovery the drawing order of handwritten fonts that less than five strokes.

Aiming at the problem that the features extracted from the color image by existing copy-move forgery detection (CMFD) algorithms are not comprehensive and the matching time is too high,the blind identification algorithm for digital images by using quantized color features and SURF detector was studied.In the feature extraction process,the algorithm combines HSV fuzzy quantization color feature and SURF feature to form a comprehensive description,called the FCQ-SURF features,of color image content.K-Means clustering and KNN method are used to improve matching efficiency in feature matching stage.The experimental results show that the algorithm can detect and locate the colorima-ge copy-move forgery well in CASIA 1.0 and FAU color image test library.It also has a good detection effect for multiple tampering attacks and multi-region tampering of images.The experimental results demonstrate that the proposed algorithm has higher detection accuracy and better matching time for color image copy-move forgery.

Batch Normalization (BN) algorithm has become a key ingredient of the standard toolkit for training deep neural networks.BN normalizes the input with the mean and variance computed over batches to mitigate the possible gradient explosion or disappearance during training of deep neural networks.However,the performance of BN algorithm often degrades when it is applied to small batch sizes due to inaccurate estimates of mean and variance.Batch ReNormalization (BRN) normalizes the input with the values of exponentialmoving average (EMA),reducing the dependency of the normalization algorithm on batches.This paper proposed a novel normalization algorithm with improved estimate on the moving mean and varianceby changing the initial value of EMA and adding corrections to the estimates.The experimental results show that the proposed algorithm has better performance in convergence speed and accuracy than both the standard BN and BRN algorithms.

The car lights not only have lighting functions,but also are important ways for vehicles to communicate with other vehicles while driving.In assisted driving,understanding the light message transmitted by surrounding vehicles accurately is a prerequisite for making correct driving decisions.During daytime driving,due to thechangeble environment,it is difficult to achieve good results in road measurement by matching the lights and then recognizing the lamp language.To this end,in view of the daytime driving situation,this paper proposed a method of light language recognition based on vehicle detection.In this paper,the Adaboost cascade classifier is trained to test the vehicle by using the training method of the updated sample.Based on this,the position distribution feature of the vehicle rear is used to determine the region of interest of the lights.In the RGB color space,a color segmentation algorithm is proposed,which can accurately extract the position of the lamp,judge the lighting state of the lamp on the basis of the region of interest,and eliminate the misdetection of color segmentation algorithm.This paper uses the brightness feature when the lamp is lit.The high-position brake light is used as the recognition condition of the brake light lamp language,and the historical frequency information is used as the recognition condition of the turn signal light,and the recognition of the front taillight light during daytime driving is completed.The experiment uses VS2010 and opencv3.4.9 as the algorithm implementation tool,and uses the actual measured data of the driving recorder provided by SAIC as the test data.After test,the accuracy of classifier recognition in experimental training is 93％.Compared with the traditional Adaboost classifier,the recognition accuracy is improved by about 2%,the average accuracy of the light recognition algorithm is 93％,and the average time of the algorithm is about 53ms.The test results show that the classification training method used in the experiment can improve the detection accuracy slightly,and the light recognition algorithm can accurately identify the brakes,the turn signals and the two kinds of lights simultaneously,and can basically guarantee the real-time performance.

Gesture recognition based on WiFi channel state information (CSI) has broad application prospects in human-computer interaction.At present,most methods require manual extraction of features,and the feature extraction process is cumbersome.It can only recognize gestures in a specific direction,which limits the range of people’s activities.To solve the above problems,this paper proposed a method based on Long Short-Term Memory (LSTM) training to design a CSI gesture recognition system based on LSTM.The system preprocesses the collected CSI data through such as abnormal point removal,optimal subcarrier selection and discrete wavelet variation denoising.The LSTM network trains the classification without manual extraction of gesture features.Finally,the recognition of four gestures is achieved,which are pushing,pulling,left swing and right swing in four different directions,and an average recognition accuracy of 82.75% is reached.This paper discussed the influence of the distance between sender and receiver and the size of the data set on the accuracy of gesture recognition,and compared the gestures in four directions by WiG and WiFinger.The results show that the proposed method has higher recognition effect.

Harris Feature point detection is widely used in target recognition,tracking and 3D reconstruction.The computation of the feature point detection algorithm for big data problem is time-consuming and computation-intensive.There is a problem of large time-consuming and low efficiency in the algorithm of feature points detection with large data quantity.In the multi-CPU programming model based on OpenMP and GPU parallel environment based on CUDA and OpenCL architecture,In this paper,the parallel algorithm of the Harris feature point detection was proposed.In the comparison experiment of hallFeng image set on different platforms,the experimental results show that the multi-CPU feature point detection algorithm based on OpenMP shows good multi-core scalability,and the parallel feature point detection algorithms based on CUDA and OpenCL architecture in GPU parallel environment can obtain high speedup and good data and platform scalability,the maximum speed up can be more than 90 times,and the acceleration effect is significant.

There are many mismatches in traditional feature point matching,and the efficiency is not high.Aiming at mismatching,this paper proposed a method of screening based on binary mutual information.According to the mutual information of feature points,the matching of feature points is judged correctly.In addition,the feature points extracted by ORB algorithm are distributed in the region of color change,which is more centralized.The transformation matrix obtained by RANSAC algorithm is only applicable to the region of feature points distribution,which makes the stitching result error.In order to solve this problem,this paper used the improved RANASC algorithm to screen out the interior points firstly,and then used the interior points to get the new feature points.In this way,feature points can be disper-sed,and the iterative method is used to get the best transformation matrix.The results show that when binary mutual information is used to screen feature points,it improves the accuracy of matching and increases the number of feature points matching.The improved RANSAC algorithm can effectively solve the problem of few and more concentrated feature points and make the result of image mosaic more accurate.

At present,the most difficult problem in continuous sign language recognition is how to split out the words effectively.In this paper,key actions were regarded as the basic units of sign language and an algorithm based on double transfer probability of key actions was proposed.After acquiring the sequence of basic units from continuous sign language,the boundaries of words can be effectively found by judging the intra-word and inter-word transfer relations of all adjacent basic units.Then the sequence of basic units are segmented by these boundaries and the candidate words of each group of basic units can be identified.Finally,according to the transfer probabilities between candidate words of different groups,the probability of corresponding synthetic sentence is calculated and then the final recognition result is output by the principle of maximum probability.The algorithm is easy to implement and has high execution efficiency.It can be applied to non-specific population through experimental verification.

Sign language recognition is an effective way to break the barriers between communication between deaf and hearing people.Generally,Chinese sign language can be divided into gesture language and finger language.Regional and individual differences lead to a wide variety,therefore gesture language recognition is relatively difficult,which requires constant learning and training.The finger language gives the result through the expression of the Chinese pinyin letters,which is deterministic,especially in terms of name,special meaning,and abstract expression.Most of the researches in sign language recognition concentrate on a certain gesture,focusing on key features such as hand shape,direction,position and motion trajectory,and combine some learning algorithms to improve the recognition accuracy,but neglect the most basic and reliable finger recognition.To this end,an effective method using gray level co-occurrence matrix (GLCM) and fine Gaussian support vector machine (FGSVM) was proposed to solve the problem of identifying Chinese finger sign language more accurately and effectively.The research method is as follows.Firstly,the finger sign language data set was constructed.The finger language image was directly obtained by the digital camera or got from the key frame of the video,meanwhile the hand shape was segmented from the image,and each image was adjusted to N×N specific size and converted to grayscale images.Secondly,feature extraction was performed to reduce the dimension of the intensity values in the grayscale image,and at the same time,the corresponding gray level co-occurrence matrix was created,and the enhanced data features were obtained by adjusting the parameters of inter-pixel distance and angle.Finally,the extracted image feature data were submitted to the fine Gaussian support vector machine classifier based on the 10-fold cross-validation classification.Experiments on 510 Chinese finger sign language image samples from 30 categories show that the classification accuracy based on GLCM-FGSVM is up to 92.7%,and this method can be considered as effective approach in Chinese finger sign language classification.

At present,deep learning is a research hotspot based on image traffic sign detection and recognition proces-sing,and has achieved remarkable results.Aiming at the problem of traffic sign detection and recognition based on car-video,this paper proposed a real-time detection and recognition algorithm for traffic signs based on Spatiotemporal-CNN (ST-CNN).It constructs a Spatiotemporal model (STM) based on the spatiotemporal relationship between frames of image sequences,and combines the STM with Convolutional Neural Network (CNN).The experimental results show that the algorithm can detect,screen,track and identify the same traffic sign in the video image sequence.It can effectively reduce CNN data input and system resource consumption,and improve computational efficiency,while ensuring high accuracy.It satisfies the real-time requirements of traffic sign detection and recognition in video.The algorithm takes an average of 26.82 milliseconds per frame and the recognition accuracy reaches 96.94%.

Multi-focus image fusion uses many complementary information of the image to obtain a clear fused image.In traditional multi-scale analysis methods,image information is easily lost due to sampling and fusion strategies.In sparse representation methods,due to the lack of dictionary expression ability,the fusion details are blurred and the fusion time complexity is very high.For multi-focus image fusion method based on spatial domain method,the algorithm for measuring image activity level is very critical.A fractional differential feature is proposed to measure the activity level of the image.The algorithm first convolves the image with a fractional mask in eight directions,and then accumulates the absolute value after convolution in each direction to obtain the activity level measurement of the original image.Each metric map is then compared separately by using a sliding window technique.The sum of the windows and the large ones is regarded as the focus,and the corresponding score map is incremented by one.The decision map is obtained by the score map information.Finally,the final fused image is obtained by weighting the original image by decision graph.Through experimental comparison and analysis,this algorithm has certain advantages over traditional algorithm.

A new spatial localization mechanism based on CT image-guided robot-assisted percutaneous lung biopsy was proposed.Firstly,six marker points are designed and fixed on the CT examination bed at the same time to reference and locate in hardware.Secondly,the improved D-H inverse motion algorithm is used in software to guide the robot to perform percutaneous lung puncture.The simulation results show that the success rate of one-time puncture can be effectively guaranteed by using the location mechanism proposed in this paper.

With the rapid development of graphic image processing technology,image enhancement and recognition methods have been widely used in various industries.On this basis,text recognition technology has also made great progress.Aiming at the problems of shui text random brush strokes,variable fonts and more noise,this paper proposed an improved image enhancement and recognition method.The median filtering algorithm is used to reduce image noise,and the histogram equalization method is used to enhance image contrast.The binarization process is executed to extract the target text in the image,and the corrosion expansion process is executed to refine and expand the background.Finally,the improved text extraction algorithm is used to highlight the outline of the shui text,and the Sobel operator is used to extract the edge of the shui text.The simulation contrast experiment was carried out.The experimental results show that the method effectively reduces image noise,and accurately extracts shui characters.The method can be used in the fields of national character extraction and recognition,cultural relics restoration,image enhancement,etc.It is of great significance for protecting the heritage of ethnic cultural relics and carrying forward the traditional culture of ethnic minorities.

The illumination of the nighttime image is uneven,the overall brightness is low,the color deviation is large,and there is halo near the artificial light source.Existing deblurring models and algorithms often remove the effects of uneven illumination by estimating the illumination map in the case of uneven illumination.By combining the deep learning method with the radial basis function neural network,the illumination intensity was extracted,and the night image deblurring algorithm based on illumination estimation was proposed.For the problem of uneven illumination,the modulation transfer function (MTF) in the imaging process is calculated by estimating the illumination map.Taking the point diffusion function of the transport degrada-tion model as prior knowledge,combining the mathematical model of semi-blind image restoration method,the target image is processed to improve the quality of night vision imaging.In addition,the effectiveness of this method is verified by comparing with the traditional blind restoration method,and the image quality is improved evidently.

In the process of transcranial magnetic stimulation treatment,it is important to accurately and quickly detect the posture of the human head.Aiming at the problem that the head pose estimation based on two-dimensional color image is sensitive to the environment and posture,a head posture detection method combining both color image and depth image was proposed.The two-dimensional position information of the face feature points is detected by the color image,and the three-dimensional head coordinate system is defined by combining the depth information;Then,based on the existing ICP point cloud registration algorithm,a coarse registration method was proposed.The initial pose parameters are obtained by calculating the transformation relationship between the coordinate system of the head cloud,to be detected and the standard head point cloud,to protect the point cloud registration from falling into local optimum.Experiments show that the algorithm can accurately detect the head posture of the human body in a consulting room environment where the light source is uniform and sufficient,and improve the robustness of the attitude estimation when the head posture angle is large.

Breast cancer is one of the most common malignant tumors of women.The incidence of breast cancer is increasing year by year,which seriously threatens the health of the patients.In recent years,more and more attention has been paid to how to replace the traditional needle biopsy in the diagnosis of benign and malignant breast nodules.Medical research shows that significant differences exist on the edge of benign and malignant nodules.So the algorithm of boundary enhancement treatment provides a new way for the study of judgment of benign and malignant breast cancer.The database was constructed with the support of Beijing Friendship Hospital which is affiliated to Capital Medical University.The images are expanded based on the comparison of 5 kinds of boundary enhancement (ACE) algorithm.AlexNet network model is used which is excellent in image classification.The data processed by linear,nonlinear contrast stretching,histogram equalization,histogram thresholding and adaptive contrast enhancement algorithm are applied to the AlexNet model.The influence of the five algorithms on the accuracy of AlexNet model is compared,and a preprocessing algorithm,which is more suitable for ultrasonic images of breast nodules,is obtained.The total number of images in the expanded data set is more than ten thousand,of which the training set is 80%,and the verification set and the test set account for 10% each.Finally,the sensitivity,specificity and accuracy parameters are calculated by plotting the ROC curve,and the test results are evaluated.The better test results are obtained.

Aiming at the problems that the existing object recognition methods are easy to be influenced by “uncontrolled” factors such as illumination,angle,size and complex environment,and have the problems such as low recognition rate,poor real-time performance and large memory consumption,this paper proposed a new object recognition algorithm,on which the object recognition system based on mobile terminal was realized.This method first employs particle filter algorithm to track the detection range by adding windows,and then applies the watershed segmentation algorithm to segment objects,then uses the HOG(Histogram of Oriented Gradient) algorithm to extract object features.Finally,the random forest algorithm is utilized to recognize objects.The experimental results show that this method can be used to identify the mobile terminal quickly and accurately in an “uncontrolled” environment.

K-means clustering algorithm is a common way in image segmentation.As an unsupervised learning method,it can find the association rules from characteristics of grey levels,thus has a great capability of segmentation.However,due to its single classification basis and uncertainty of the initial cluster centers,this algorithm still has some defects in image segmentation.Aiming at this problem,this paper proposed a modified K-means algorithm for image segmentation.The new algorithm uses the improved iterative algorithm based on information entropy to select thresholds as the initial K-means clustering centers,and then puts forward a new weighted quality evaluation function for K-means algorithm to get better segmentation thresholds.The experimental results show that the improved algorithm has higher accuracy and stability than OTSU algorithm and traditional K-means algorithm in image segmentation.

Obtaining depth information from 2D images is a hot topic in the field of computer vision.Classical binocular vision methods require camera parameters and multiple images of the same scene.Insufficient visual parameters can easily lead to errors in calculation,while a single image can only rely on its own geometric information to get the image depth.This paper used the geometric information of the image and the binocular vision model to get the depth value of the object in a single ordinary two-dimensional image with unknown camera parameters.The experimental results show that the image depth obtained by the proposed method can relatively accurately reflect the real information of the scene,which is consistent with the actual observation results.

Static gesture recognition has caught special attention for its great application value in man-machine interaction.At the same time,the accuracy of gesture recognition is affected by the complexity of gesture background and the diversity of gesture morphology in a certain extent.In order to improve the accuracy of gesture recognition,a method was proposed,which is based on convolutional neural network(CNN) and random forest(RF).Firstly,the image of the static gesture is segmented,then the feature extraction function of convolution network is used to extract feature vectors,and finally the random forest classifier is used to classify these feature vectors.On the one hand,the CNN has the ability of layered learning and is able to collect more representative information on the picture.On the other hand,random forest shows randomness for samples and feature selection,meanwhile,it can be avoided easily that the results of each decision tree is averaged over fitting problem.This paper verified by using the static gesture data set,and the experimental results show that the proposed method can effectively identify the static gestures and achieve an average recognition rate of 94.56%.The method proposed in this paper was further compared with principal component analysis(PCA) and partial binary(LBP).The experimental results show that the classification and recognition effect with feature extraction by CNN is better than PCA and LBP.The recognition rate is 2.44% higher than that of PCA-RF methodand 1.74% higher than that of LBP-RF method.Finally,the recognition rate of the proposed method reaches 97.9%,which is higher than the other two traditional feature extraction methods.

It is difficult to maintain the image texture details in image interpolation technology.To overcome this problem,this paper proposed a new method of image interpolation based on rational interpolation function.Firstly,image is automatically divided into texture regions and smooth regions using the isoline method.Secondly,a new type of C²-continuous rational interpolation function is constructed,which is an organic unity of polynomial models and rational mo-dels.According to regional features of the image,the texture region is interpolated by rational model and the smooth region is interpolated by polynomial model.Finally,based on the human visual system,this paper proposed a multi-scale approach to boost details of interpolated image.Experimental results show that this algorithm not only has lower time complexity,but also can preserve image detail,and obtain high objective evaluation data.

With the in-depth development of underwater work in rivers,lakes and offshore near-shore shallow water areas,diver’s underwater engineering construction such as underwater salvage,positioning and exploration becomes significant.The TKIS-I helmet-mounted colorful imaging sonar developed by this lab has been acknowledged by Navigation and Warranty Department of Chinese Navy.Currently,there are more than two dozens of TKIS-I in service.However,under the complex underwater environment,divers usually perform underwater operations with great risks,so it is expected to use underwater robots to achieve automatic underwater target detection in the future.Aiming at the feature of sonar image,this paper adopted feature extraction methods of HSV color space,Histogram of Oriented Gradient(HOG) and Local Binary Pattern(LBP) respectively in the aspects of color,shape and texture.Besides,the paper improved multi-feature fusion method and used optimized support vector machine(SVM) for classification,aiming to quickly detect underwater targets to lay the foundation for robots’ underwater automatic target detection in the future.

Seat belts are one of the most effective measures to protect safety of drivers which the law stipulates that drivers must wear seat belts when driving the vehicle.At present,the identification of seat belt during driving is mainly based on manual screening.However,the traditional detection methods can not meet the needs of traffic management as the rapid increase of the number of vehicles.And the automatic processing of seat belt detection has become one of the urgent problems in the current traffic system.In this paper,a recognition system for seat belts of drivers is designed.First,the vehicle window is roughly positioned by the geometric relationship between the license plate and the window.Second,Hough transform is used to detect the upper and lower edges of the window and the integral projection transformation is used to detect the left and right borders of the window.The detected pictures will be cut into half to get the driver rough position.Finally,the seat belt identification analysis based on deep convolutional neural network is conducted which adds spatial transform layer.Experiments are carried out on different bayonet and different time periods for 10000 pictures.The experimental results show that the proposed method can effectively identify whether the driver wears the seat belt according to the regulations,and the comprehensive recognition rate is significantly improved compared with the existing method.

How to make full use of the information contained in the image for super-resolution reconstruction is still an open question.This paper proposed an image super-resolution reconstruction algorithm based on adaptive sparse representation and non-local self-similarity.In the process of training and reconstruction,the K-means algorithm is used to cluster the selected datasets,and similar image blocks are gathered together.Then PCA is used to process the adaptive selection dictionary for super-resolution reconstruction.Compared with image reconstruction through a fixed dictionary,the adaptive selection dictionary is used to reconstruct the image,and the effect of reconstructed image obtained will be more superior.The experimental results on natural images show that the super-resolution images reconstructed by the proposed algorithm are more detailed,with fewer artifacts and sharper edges.