Due to optical aberrations in imaging optics,the image taken from simple lensessuffers from severe artifacts and blurring.Aiming at this kind of blurring problem,this paper proposed a deblurring method combining adaptive gradient sparsity and interchannel correlation.This method restores every color channel of the blurred images separately through imposing different sparse priors on points in smooth areas and at edges and using interchannel correlation constraint,which uses edge information preserved in some channel to restore another channel.The simulation experiment results show that the proposed method can achieve better restoration in respect of image resolution and visual effect for blurred images through simple lens.

Billions of Wi-Fi assess points (APs) have been deployed to provide wireless connection to people with different kinds of mobile devices.Toaccelerate the speed of Wi-Fi connection,mobile devices will send probe requests to discover nearby Wi-Fi APs,and maintain previously connected network lists (PNLs) of APs.Previous studies show that the Wi-Fi fingerprints that consist of probed SSIDs individually will leak private information of users.This paper investigated the privacy caused by the Wi-Fi fingerprints in the wild,and provided a data-driven solution to protect privacy.First,measurement studies were carried out based on 27 million users associating with 4 million Wi-Fi APs in 4 cities,and it was revealed that Wi-Fi fingerprints can be used to identify users in a wide range of Wi-Fi scenarios.Based on semantic mining and analysis of SSIDs in Wi-Fi fingerprints,this paper further inferred demographic information of identified users (e.g.,people’s jobs),telling “who they are”.Second,this paper proposed a collaborative filtering (CF) based heuristic protection method,which can “blur” an user’s PNL by adding faked SSIDs,such that nearby users’ PNLs and Wi-Fi fingerprints are similar to each other.Finally,the effectiveness of the design was verified by using real-world Wi-Fi connection traces.The experiments show that the refined PNLs protect users’ privacy while still provide fast Wi-Fi reconnection.

In recent years,the rapid growth of demand based on location-based services has led to the development of positioning technology.The vision-based approach utilizes multiple images to restore more accurate camera pose para-meters,but there is no uniform assessment of the performance of its quantitative evaluation.Now the mainstream camerapose assessment method is compared with the GPS data.However,since the photo comes with the GPS tag noise and the conversion between different coordinate systems introduces errors,using GPS tag as ground truth to evaluate the accuracy of the estimated camera pose is not an objective way.In this paper,an objective accuracy evaluation method was proposed.The reference plane was established by the calculated pose.The camera pose obtained by the PnP algorithm was projected onto the reference plane by the same method.

The purpose of this paper is to accurately estimate the crowd density in real scenes based on image information from arbitrary perspective and arbitrary crowd density.However,crowd counting on static images is a challenging problem.Due to the perspective distortion and the crowd crushes caused by the projection from 3D space into 2D space,it is difficult to distinguish the difference between individual and individual and the difference between individual and background.To this end,this paper proposed a flexible and efficient scalable modularized convolutional neural network (CNN) architecture.The network allows to directly input images with arbitrary size and resolution and it does not require additional computational changes in view information.Each module of the architecture employs a multiple column structure with different convolution kernels,which can be used to fit individual information of different distances.The proposed module also combines the feature information of the front and rear two layers,reducing the decrease loss of the accuracy caused by the vanishing of the gradient.Experiments show thatthe accuracy of proposed method is increased by 14.58% and 40.53%,and the root mean square error is reduced by 23.89% and 33.90% respectively on ShanghaiTech PartA and PartB datasets compared with the state-of-the-art MCNN methods.

Human action recognition is an important research direction in the field of computer vision and pattern recognition.The complexity of human behavior and the variety of action performing make behavior recognition still as a challenging subject.With the new generation of sensing technology,RGB-D cameras can simultaneously record RGB images,depth images,and extract skeleton information from depth images in real time.How to take advantages of above information has become the new hotspot and breakthrough point of behavior recognition research.This paper presented a new feature extraction method based on Gaussian weighted pyramid histograms of orientation gradients for RGB images,and built an action recognition framework fusing multiple features.The feature extraction method and the framework proposed in this paper were researched on three databases:UTKinect-Action3D,MSR-Action 3D and Florence 3D Actions.The results indicate that the proposed action recognition framework achieves the accuracy of 97.5%,93.1%,91.7% respectively.It shows the effectiveness of the proposed action recognition framework.

Image co-segmentation is one of the active research areas in computer vision.The ability to utilize the information of similar objects in segmentation process is one of the advantages of co-segmentation,which is different from other segmentation methods.Meanwhile,establishing the similarity of corresponding objects is becoming a challenging task.This paper presented a novel consistency of center sensitive histogram for image co-segmentation.Unlike the traditional image histogram that calculates the frequency of occurrence for the intensity value by adding ones to the corresponding bin,a consistency of center sensitive histogram is computed at each pixel and a floating-point value is added to the corresponding bin for each occurrence of the intensity value.The floating-point value is a spatial consistency between the pixel of occurrence of intensity and the pixel where the histogram is computed.Therefore,the histogram not only takes the distribution of each pixel’s intensity value into account,but also the spatial relative position.A robust co-segmentation framework was proposed.Its robustness reflectsthe similar objects under different illumination and deformation condition can be both segmented well.The proposed technique was verified on various test image data sets.The experimental results demonstrate that the proposed method outperforms the average of state-of-the-art 3%,especially when the test image is in different illumination conditions and has different shapes.

Motion blur is caused by the relative motion between object and imaging system,and precise motion blur parameters are needed when the uniform linear motion-blurred image is recovered.It can be proved that the motion blur parameters are relative to the zero points of Fourier spectrum.The number of dark lines is related to the fuzzy scale,and the spectrum dark lines are perpendicular to the angle.In the detection of the dark lines of spectrum,due to the image structure or noise,it is difficult to accurately locate the spectral dark lines,and the spectral structure will be affected by image’s aspect ratio.To deal with these problems,this paper proposed a way named bottom-hat which are used in morphology and processed it in the blur image spectrum,and then used the Hough transform to find the blurred angles.Blurred angles and the distance between two mid zero points were used to find the blurred length.Experimental results show that the errors of blurred length detected by the proposed method are lower than 0.25 pixels,and the errors of the angle are lower than 0.6°.The proposed method is very robust for it can estimate the blur parameters of blurred images in different scales and contents.

With the further research on age feature extraction and age feature classification pattern,in order to make further efforts to meet the application demand of human-computer interaction system based on age information in real life,constructing an effective machine learning algorithm has become a research focus in age estimation technology of face image.Firstly,this paper analyzed the rule of multiple regional features changing with age,and divided the face into prefrontal region,eye region,central region and integrated region.Then,it constructed features extraction model of deep convolutional neural network models separately to extract age features of each region.Thirdly,taking Morph face database as the sample set,this paper divided it into 6 stages aged 10～19,20～29,30～39,40～49,50～59,and 60 years or older to train and test age feature extraction network model in multiple regions.Finally,according to the accuracy of age feature classification model,this paper defined the region-based dynamic weight age estimation model.The experiment shows that the accuracy of age estimation on Morph face database is 72.6%,and the age classification category has been raised from 4 to 6.

Video caption detection and extraction is one of the key technologies forvideo understanding.This paper proposed a two-stage approach which divides the process into caption frame and caption area,improving the caption detection efficiency and accuracy.In the first stage,caption frame detection and extraction is conducted.Firstly,the motion detection is performed according to the gray correlation frame difference,the captions are judged initially,and a new binary image sequence is obtained.Then,according to dynamic characteristics of ordinary captions and scrolling captions,the new sequence is screened two times to get caption frame.In the second stage,caption area detection and extraction is conducted.Firstly,the Sobel edge detection algorithm is used to detect the caption region,and the background is eliminated according to the constraint height.Then according to the aspect ratio,the vertical and horizontal captions are distinguished,and all captions in the caption frame can be obtained,including static captions,ordinary captions and scrol-ling captions.This method reduces the frames which need to be detected and improves caption detection efficiency by 11%.The experimental results show that the proposed method can approximately improve the F score by 9% compared with the methods of separately using the gray correlation frame difference and edge detection.

Aiming at the disadvantages of shuffled frog leaping algorithm (SFLA),such as high computational comple-xity and poor optimization efficiency,an improved shuffled frog leaping algorithm (ISFLA) was proposed in this paper.The following improvements have been made on the basis of SFLA.Firstly,the method which only updates the worst frog in SFLA is replaced by the method which updates all frogs in each group.This replacement can increase the probability of obtaining the high quality solutions,omit the steps of setting the number of iterations in the group and then improve the optimization efficiency and operability.Secondly,the method based on local optimum updating and the method based on global optimum updating are combined into a hybrid disturbance updating method,which avoids the tedious condition selection steps and further improves the optimization efficiency.Finally,the random updating method is removed to avoid destroying the superior solutions and further enhance the overall performance optimization.ISFLA was tested on the benchmark functions from CEC2005 and CEC2015,and was applied to the multi-threshold gray and color images segmentation based on Renyi entropy.The experimental results show that,ISFLA obtains higher optimization efficiency and is more suitable for threshold selection of multi-threshold image segmentation compared with SFLA and the state-of-the-art LSFLA.

This paper proposed a 2D-to-3D conversion algorithm for badminton video.The most attractive part of badminton video is the foreground.The core of the depth map extraction is to separate the foreground objects accurately from the background.The improved grab cut segmentation algorithm is used to extract foreground regions.A model for the background depth is constructed based on the structure of scene.The depth value is assigned for foreground based on the distance of scene objects from the viewpoint and the background depth map.Then the depth of foreground and background are merged.Finally,the synthesized stereo pairs of images for 3D display are obtained by DIBR formula.The experimental results show that the generated stereo images have good 3D perception performance.

Quality evaluation of color image is of great significance in image acquisition,compression,storage,transmission and so on.However,traditional objective evaluation methods often lose some color information or ignore the integrity of color image,making the results can not be well consistent with the subjective scores.This paper proposed an objective quality evaluation method of color image based on discrete quaternion Fourier transform(DQFT).A color image is expressed as a quaternion matrix and the discrete quaternion Fourier transform is applied.Then,the spectrum is divided into non-uniform bins and a reduced space representation of the spectrum is obtained by considering the characte-ristics of Human vision system which is sensitive to the distortion of lower frequency components and insensitive to higher frequency components.Next,the amplitude similarity and phase similarity between the distorted image and the reference image are described.Taking into account the influence on the image quality of the amplitude similarity and phase similarity,both of them are integrated by using entropy method and the index of the distorted image quality is achieved.Finally,image databases were used to analyze the correlation between the objective scores and the subjective scores.The experimental results demonstrate the feasibility and effectiveness of the proposed method.