In order to remove the noise in Low Dose Computed Tomography （LDCT） images and enhance the display effect of the denoised images， an LDCT image enhancement algorithm based on Generative Adversarial Network （GAN） was proposed. Firstly， GAN was combined with perceptual loss and structure loss to denoise the LDCT image. Then， dynamic gray?scale enhancement and edge contour enhancement were performed to the denoised image respectively. Finally， Non?Subsampled Contourlet Transform （NSCT） was used to decompose the enhanced image into multi?directional coefficient sub?images in the frequency domain， and the paired high? and low?frequency sub?images were adaptively fused with Convolutional Neural Network （CNN） to reconstruct the enhanced Computed Tomography （CT） image. Using the real clinical data of the AAPM competition as the experimental dataset， the image denoising， enhancement， and fusion experiments were carried out. The results of the proposed method are 33.015 5 dB， 0.918 5， and 5.99 on Peak Signal?to?Noise Ratio （PSNR）， Structural Similarity Index Measure （SSIM）， and Root Mean Square Error （RMSE） respectively. Experimental results show that the proposed algorithm retains the detailed information of the CT image while removing noise， and improves the brightness and contrast of the image， which helps doctors analyze the patient’s condition more accurately.

Traditional 2D convolutional models suffer from loss of semantic information and lack of sequential feature expression ability in sentiment classification. Aiming at these problems, a hybrid model based on 1D Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) was proposed. Firstly, 2D convolution was replaced by 1D convolution to retain richer local semantic features. Then, a pooling layer was used to reduce data dimension and the output was put into the recurrent neural network layer to extract sequential information between the features. Finally, softmax layer was used to realize the sentiment classification. The experimental results on multiple standard English datasets show that the proposed model has 1-3 percentage points improvement in classification accuracy compared with traditional statistical method and end-to-end deep learning method. Analysis of each component of network verifies the value of introduction of 1D convolution and recurrent neural network for better classification accuracy.

To improve the face detection accuracy of mobile devices, a new real-time face detection algorithm for mobile devices was proposed. The improved Viola-Jones was used for a quick region segmentation to improve segmentation precision without decreasing segmentation speed. At the same time, the optical flow estimation method was used to propagate the features of discrete keyframes extracted by the sub-network of a convolution neural network to other non-keyframes, which increased the efficiency of convolution neural network. Experiments were conducted on YouTube video face database, a self-built one-minute face video database of 20 people and the real test items at different resolutions. The results show that the running speed is between 2.35 frames per second and 22.25 frames per second, reaching the average face detection level; the recall rate of face detection is increased from 65.93% to 82.5%-90.8% at rate of 10% false alarm, approaching the detection accuracy of convolution neural network, which satisfies the speed and accuracy requirements for real-time face detection of mobile devices.

Since the mean Average Precision (mAP) of Single Shot multibox Detector (SSD) drops significantly when evaluating with higher Intersection over Union (IoU), a feature aggregation method using transposed convolution as main component was proposed. On the basis of SSD model, a deep Residual convolutional Network (ResNet) with 101 layers was used to extract features. Firstly, abstraction of semantics and context information was generated by using transposed convolutional layers which doubled the scales of deeper feature maps. Secondly, fully connected convolutional layers were applied to shallow layers to prevent unexpected bias. Finally, the shallow and deep feature maps were concatenated together, and convolutional layers with kernel size 1 were used to reduce the channel sizes. The feature aggregation can repeat multiple times. The experiments were conducted on KITTI dataset and took 0.7 as IoU threshold. Experimental results show that the mAP was improved by about 5.1 and 2 percent points compared to the original SSD model and the state-of-the-art Faster R-CNN model. The feature aggregation model can effectively improve the mAP and generate high quality bounding boxes in object detection tasks.

For smoke recognition problem, the traditional recognition methods based on sensor and image feature are easily affected by the external environment, which would lead to low recognition precision if the flame scene and type change. The recognition method based on deep learning requires a large amount of data, so the model recognition ability is weak when the smoke data is missing or the data source is restricted. To overcome these drawbacks, a new smoke recognition method based on deep transfer learning was proposed. The main idea was to conduct smoke feature transfer by means of VGG-16 (Visual Geometry Group) model with setting ImageNet dataset as source data. Firstly, all image data were pre-processed, including random rotation, cut and overturn, etc. Secondly, VGG-16 network was introduced to transfer the features in the convolutional layers, and to connect the fully connected layers network pre-trained by smoke data. Finally, the smoke recognition model was achieved. Experiments were conducted on open datasets and real-world smoke images. The experimental results show that the accuracy of the proposed method is higher than those of current smoke image recognition methods, and the accuracy is more than 96%.

To solve the problem that classical Mutual Information (MI) image registration may lead to local extremum, a registration method for multispectral magnetic resonance images based on Cross Cumulative Residual Entropy (CCRE) was proposed. Firstly, the gray level of reference and floating images were compressed into 5 and 7 bits. Then the Hanning windowed Sinc interpolation was used to calculate the CCRE of 5-bit grayscale images, and the Brent algorithm was used to search the CCRE to get the initial transformation parameters of pre-registration. Finally, the Partial Volume (PV) interpolation was adopted to calculate the CCRE of 7-bit grayscale images, and the Powell algorithm was applied to optimize the CCRE to get final parameters from the pre-registration parameters. The experimental results show that the robustness of the proposed method is improved compared with the CCRE registration of PV interpolation, while the registration time is saved about 90% and accuracy is improved compared with the CCRE of Hanning windowed Sinc interpolation. The presented method ensures robustness, efficiency and accuracy, so it is suitable for multi-spectral image registration.

In the simulation of Software Defined Network (SDN), the existing network simulation tools usually do not consider the processing delay of SDN switchs. To make the simulation result more realistic and accurate, a scheme to simulate the processing delay was proposed. First, the scheme divided the process of the switch forwarding into two aspects: inquiry operations on flow table and execution of various actions, and then transferred the two aspects into processing delay by using processor frequency and memory cycle. Measurement and comparison were conducted on the processing delay of switches with different configuration in real and simulation environments. The results show that the simulated processing delay of the proposed method is almost close to that in real environment, it can accurately estimate the processing delay of switches.

To solve the problem of Fine Particulate Matter (PM2.5) concentration prediction, a PM2.5 concentration prediction model was proposed. First, through introducing the comprehensive meteorological index, the factors of wind, humidity, temperature were comprehensively considered; then the feature vector was conducted by combining the actual concentration of SO2, NO2, CO and PM10; finally the Least Squares Support Vector Machine (LS-SVM) prediction model was built based on feature vector and PM2.5 concentration data. The experimental results using the data from the city A and city B environmental monitoring centers in 2013 show that, the forecast accuracy is improved after the introduction of a comprehensive weather index, error is reduced by nearly 30%. The proposed model can more accurately predict the PM2.5 concentration and it has a high generalization ability. Furthermore, the author analyzed the relationship between PM2.5 concentration and the rate of hospitalization, hospital outpatient service amount, and found a high correlation between them.

Since the isotropic diffusion will easily blur edge features,and coherence-enhancing diffusion will produce pseudo striations in the background regions during the denoising process, a weighted diffusion algorithm was proposed to reduce the Rician noise of Magnetic Resonance Imaging (MRI) image according to the distribution of noise. A threshold value was calculated by the Rician noise variance in the background region of MRI image, which might be used to distinguish the image background and the edge of Region-Of-Interest (ROI). A weighting function combining the isotropic diffusion and the coherence-enhancing diffusion based on the calculated value was constructed. The constructed function could adaptively adjust the weight values of two kinds of diffusion in different structural regions in order to give full play to the advantages while overcoming the disadvantages of the above two kinds of diffusion.The experimental results show that it is better than some classical diffusion algorithms in Peak Signal-to-Noise Ratio (PSNR) and Mean Structural Similarity(MSSIM).Thus, it has better performance on noise reduction and edge preservation or enhancement.

Multiple Sclerosis (MS) is a chronic disease that affects the central nervous system and MS lesions are visible in conventional Magnetic Resonance Imaging (cMRI). A new method for the automatic segmentation of MS White Matter Lesions (WML) on cMRI was presented, which enabled the efficient processing of images. Firstly the Kernel Fuzzy C-Means (KFCM) clustering was applied to the preprocessed T1-weight (T1-w) image for extracting the white matter image. Then region growing algorithm was applied to the white matter image to make a binary mask. This binary mask was then superimposed on the corresponding T2-weight (T2-w) image to yield a masked image only containing white matter, lesions and background. The KFCM was reapplied to the masked image to obtain WML. The testing results show that the proposed method is able to segment WML on simulated images of low noise quickly and effectively. The average Dice similarity coefficient of segmentation result is above 80%.

The exiting model representation cannot meet the requirements of multistage modeling, so model share, reuse and management can hardly achieve in multistage modeling process. Therefore, a hierarchical model management framework based on universal relation model presentation was presented. Firstly, the requirements of model representation in model management and the limitations of exiting model representation were analyzed, then a model representation based on universal relation was investigated to set the mapping relation between layers of conceptual model and mathematical model, and the integrative model representation and the hierarchical model management framework including conceptual model, mathematical model and physical model were set up. At last, the logic of modelbase in management framework was designed and the physical model generation based on universal relation was investigated. The model of different modeling process could be managed uniformly based on an integrative model representation.

In view of complicated and inefficient calculation of system matrix, a simple length weighted algorithm was proposed. Compared with the traditional length weighted algorithm, the proposed algorithm reduced situations of the intercepted photon rays with the grid and the grid index of the proposed approach was determined in the two-dimensional coordinate. The computational process of the system matrix was improved based on the proposed algorithm. The image reconstructed with the system matrix was constructed through the new process, and the quality of the reconstructed image was assessed. The experimental results show that the operation speed of the proposed algorithm is more than three times faster than Siddon improved algorithm, and the more lengths in the length weighted algorithm get considered, the better quality of the reconstructed image has.

To solve the problem that GrabCut algorithm is sensitive to local noise, time consuming and edge extraction is not ideal, the paper put forward a new algorithm of improving image segmentation based on GrabCut. Multi-scale watershed was used for gradient image smoothing and denoising. Watershed operation was proposed again for the new gradient image, which not only enhanced image edge points, but also reduced the computation cost of the subsequent processing. Then the entropy penalty factor was used to optimize the segmentation energy function to prevent target information loss. The experimental results show that the error rate of the proposed algorithm is reduced, Kappa coefficient is increased and the efficiency is improved compared with the traditional algorithm. In addition, the edge extraction is more complete and smooth. The improved algorithm is applicable to different types of image segmentation.

The three-dimensional spatial target localization based on wireless communication and networking technology is a hot research topic in the field of Internet of Things.However,there are still some problems including location is not accurate enough,the calculation overhead is too high and power consumption is too large.Thus,a distributed three-dimensional localization mechanism for the environment of the Internet of Things is proposed.The algorithm use Cooperative Location-Sensing(CLS) to mesh grid and determine the target location by the estimated distance.It combines Gaussian fitting ,signal sorting mechanism and Bounding-inbox algorithm which have effectively reduced signal interference.Moveover,it use local meshing to reduce the grid voting overhead.Simulation results show that the algorithm is better than the existing three-dimensional localization algorithm in positioning accuracy and have a lower power consumption.

In order to improve the accuracy and recognition speed of the remote sensing image classification, this paper put forward a new algorithm of remote sensing image classification based on K-type Support Vector Machine （SVM）,and this algorithm used texture features extracted by gray level co-occurrence matrix combined with the spectral ones for classification. The classification simulation tests were done with two groups of Landsat ETM+data. The results show that the new algorithm can improve the accuracy and efficiency of the classification, raise generalization ability, and K-type SVM is a superior classifier to the Radial Basis Function (RBF) SVM.

To obtain more image details, a remote sensing image fusion algorithm based on edge features was proposed. It applied wavelet packet transform, and determined fusion rules according to the wavelet coefficients and edge features of three sub-direction bands. The experiments on TM multi-spectral image and SPOT high-resolution image were carried out, and the fusion results were analyzed from both subjective and objective perspectives. The simulation results show that this method has certain improvement in both performing dimensional detail information and maintaining spectral information.

To improve the performance of a text-independent speaker recognition system, the authors proposed a speaker recognition system based on linear log-likelihood kernel function. The linear log-likelihood kernel compressed the input cepstrum feature sequence of a speaker model by a Gaussian mixture model. The log-likelihood between two utterances was simplified to the distance between the parameters of Gaussian mixture model. Polarization identity was applied to obtain the mapping from a cepstrum feature sequence to a high dimension vector. Support Vector Machine (SVM) was used to train speaker models. The experimental results on National Institute of Standard and Technology show that the proposed kernel has excellent performance.

Spatial index is a key issue in massive spatial data processing. This paper improved the multi-layers grid by analyzing the grid files. Some creative theories and relevant algorithms were put forward such as first layer grid auto-selection algorithm based on normal distribute and new grid-contain algorithm. This paper analyzed the performance of the improved multi-layer grid spatial index by real data test. Test results show that in most case the creative theories improve the performance and adjustability of index.

The application and special requirements of time synchronization algorithm in wireless sensor network were introduced firstly. Then the un-synchronized reason, the relationship of local clock and logical clock, and the facts that impact the clock synchronizing precision were analyzed. Some promising methods deserving further study were proposed based on the classification of the existing clock synchronization algorithms.