Aiming at the problem that current detectors based on deep learning cannot effectively detect objects with irregular shapes or large differences between length and width, based on the traditional Faster Region-based Convolutional Neural Network (Faster R-CNN) algorithm, an improved two-stage object detection framework named Accurate R-CNN was proposed. First of all, a novel Intersection over Union (IoU) metric-Effective Intersection over Union (EIoU) was proposed to reduce the proportion of redundant bounding boxes in the training data by using the centrality weight. Then, a context related Feature Reassignment Module (FRM) was proposed to re-encode the features by the remote dependency and local context information of objects, so as to make up for the loss of shape information in the pooling process. Experimental results show that on the Microsoft Common Objects in COntext (MS COCO) dataset, for the bounding box detection task, when using Residual Networks (ResNets) with two different depths of 50 and 101 as the backbone networks, Accurate R-CNN has the Average Precision (AP) improvements of 1.7 percentage points and 1.1 percentage points respectively compared to the baseline model Faster R-CNN, which are significantly than those of the detectors based on mask with the same backbone networks. After adding mask branch, for the instance segmentation task, when ResNets with two different depths are used as the backbone networks, the mask Average Precisions of Accurate R-CNN are increased by 1.2 percentage points and 1.1 percentage points respectively compared with Mask Region-based Convolutional Neural Network (Mask R-CNN). The research results illustrate that compared to the baseline model, Accurate R-CNN achieves better performance on different datasets and different tasks.

Dehazing algorithms based on image training pairs are difficult to deal with the problems of insufficient training sample pairs in remote sensing images, and have the model with weak generalization ability, therefore, a remote sensing image dehazing method based on cascaded Generative Adversarial Network (GAN) was proposed. In order to solve the missing of paired remote sensing datasets, U-Net GAN (UGAN) learning haze generation and Pixel Attention GAN (PAGAN) learning dehazing were proposed. In the proposed method, UGAN was used to learn how to add haze to the haze-free remote sensing images with the details of the images retained by using unpaired clear and haze image sets, and then was used to guide the PAGAN to learn how to correctly dehazing such images. To reduce the discrepancy between the synthetic haze remote sensing images and the dehazing remote sensing images, the self-attention mechanism was added to PAGAN. By the generator, the high-resolution detail features were generated by using cues from all feature locations in the low-resolution image. By the discriminator, the detail features in distant parts of the images were checked whether they are consistent with each other. Compared with the dehazing methods such as Feature Fusion Attention Network (FFANet), Gated Context Aggregation Network (GCANet) and Dark Channel Prior (DCP), this cascaded GAN method does not require a large number of paired data to train the network repeatedly. Experimental results show this method can remove haze and thin cloud effectively, and is better than the comparison methods on both visual effect and quantitative indices.

In order to solve the problem that the clustering centers cannot be determined automatically by Density Peaks Clustering (DPC) algorithm, and the clustering center points and the non-clustering center points are not obvious enough in the decision graph, Comparative density Peaks Clustering algorithm with Automatic determination of clustering center (ACPC) was designed. Firstly, the distance parameter was replaced by the distance comparison quantity, so that the potential clustering centers were more obvious in the decision graph. Then, the 2D interval estimation method was used to perform the automatic selection of clustering centers, so as to realize the automation of clustering process. Experimental results show that the ACPC algorithm has better clustering effect on four synthetic datasets; and the comparison of the Accuracy indicator on real datasets shows that on the dataset Iris, the clustering accuracy of ACPC can reach 94%, which is 27.3% higher than that of the traditional DPC algorithm, and the problem of selecting clustering centers interactively is solved by ACPC.

Concerning the increasingly high real-time processing requirement of space science satellite observed data, a real-time processing method of space science satellite data based on stream computing framework was proposed. Firstly, the data stream was abstractly analyzed according to the data processing characteristics of space science satellite. Then, the input and output data structures of each processing unit were redefined. Finally, the parallel data stream processing structure was designed based on the stream computing framework Storm to meet the requirements of parallel processing and distributed computing of large-scale data. The developed system for space science satellite data processing applying with this method was tested and analyzed. The results show that the data processing time is half of that of the original system under same conditions and the data localization strategy has higher throughput than round-robin strategy with the data tuple throughput increased by 29% on average. It can be seen that the use of stream computing framework can greatly shorten the data processing delay and improve the real-time performance of the space science satellite data processing system.

Hypertension is an important hazard to health. Blood pressure prediction is of great importance to avoid grave consequences caused by sudden increase of blood pressure. Based on traditional Long Short-Term Memory (LSTM) network, a multi-factor cue LSTM model for both short-term prediction (predicting blood pressure for the next day) and long-term prediction (predicting blood pressure for the next several days) was proposed to provide early warning of undesirable change of blood pressure. Multi-factor cues used in blood pressure prediction model included time series data cues (e.g. heart rate) and contextual information cues (e.g. age, BMI (Body Mass Index), gender, temperature).The change characteristics of time series data and data features of other associated attributes were extracted in the blood pressure prediction. Environment factor was firstly considered in blood pressure prediction and multi-task learning method was used to help the model to capture the relation between data and improve the generalization ability of the model. The experimental results show that compared with traditional LSTM model and the LSTM with Contextual Layer (LSTM-CL) model, the proposed model decreases prediction error and prediction bias by 2.5%, 3.8% and 1.9%, 3.2% respectively for diastolic blood pressure, and reduces prediction error and prediction bias by 0.2%, 0.1% and 0.6%, 0.3% respectively for systolic blood pressure.

Because of the uncertainty and complexity of remote sensing image data, it is difficult for traditional unsupervised algorithms to create an accurate classification model for them. Pattern recognition methods based on fuzzy set theory can express the fuzziness of data effectively. In these methods, type-2 fuzzy set can better describe inter-class hybrid uncertainty. Furthermore, semi-supervised method can use prior knowledge to deal with the generalization problem of algorithm to data. Therefore, a remote sensing image classification method based on Semi-Supervised Adaptive Interval Type-2 Fuzzy C-Means (SS-AIT2FCM) was proposed. Firstly, by integrating the semi-supervised and evolution theory, a novel fuzzy weight index selection method was proposed to improve the robustness and generalization of the adaptive interval type-2 fuzzy C-means clustering algorithm. The proposed algorithm was more suitable for the classification of remote sensing data with severe spectral aliasing, large coverage areas and abundant features. In addition, by performing soft constrained supervision on small number of labeled samples, the iterative process of the algorithm was optimized and guided, and the greatest expression of the data was obtained. In the experiments, SPOT5 multi-spectral remote sensing image data of the Summer Palace in Beijing and Landsat TM multi-spectral remote sensing image data of the Hengqin Island in Guangdong were used to compare the results of the existing fuzzy classification algorithms and SS-AIT2FCM. The experimental results show that the proposed method obtains more accurate classification and clearer boundaries of classes, and has good data generalization ability.

Video transcoding is applied in the field of Internet video coding. When the original video is transcoded multiple times, only the distortion between the input video and the output video can be calculated and the distortion between the output video and the original video can not be learned. Here an algorithm for estimating the distortion between the output video and the original video was proposed to control the quality of the output program. Firstly, the superposition of distortion caused by multiple lossy transcoding was analyzed to derive the lower limit of total distortion. Then the probability method was exploited to make an estimation on the distortion between the original video and the final output video. Finally, the least square fitting was used to correct the estimation according to the prediction error. Experimental results demonstrate that the proposed algorithm can accurately estimate the distortion with the prediction error of 0.02dB, 0.05dB and 0.06dB for Y, U and V components on average respectively after correction.

Since the visually impaired person could not enjoy videos and other digital media thoroughly, in order to extend tactile perception channels for video media, an electrostatic force tactile rendering method for video perception was put forward. Firstly, target pixels of the current video frame were acquired according to the location of the finger, and color information of target pixels were transformed from RGB color model to HSI color model. Then the hue parameter of target pixels was used to map stimuli frequencies of electrostatic force, the intensity and saturation parameters of target pixels were used to map stimuli amplitudes of electrostatic force, and the tactile stimuli signal was composited to render the real-time video. Finally, dynamic color perception experiments and identification of brightness perception experiments were designed. The results show that the proposed method can realize to sense the information of objects in the video, the average accuracy of dynamic identification reaches 90.6%, the average accuracy of color identification reaches 69.4%, and the average accuracy of brightness identification reaches 80.0%. The proposed method can extract the dynamic characteristics of video information effectively and enhance the real-time tactile rendering for video.

In view of the problems that the traditional general graph matching search is inefficient, and refractive index data cannot be positioned fast in large data environment, a distributed massive molecular retrieval model based on consistent Hash function was established. Combined with the characteristics of molecular storage structures, to improve retrieval efficiency of molecules, the continuous refractive index was discretized by fixed width algorithm to establish high-speed Hash index, and the distributed massive retrieval system was realized. The size of dataset was effectively reduced, and Hash collision was handled according to the visiting frequency. The experimental results show that, in the chemical data containing 200 thousand structures of molecules, the average time of this method is about five percent of the traditional general graph matching search. Besides, the model has the steady performance with high scalability. It is applicable to retrieve high-frequency molecules in accordance with refractive index under the environment of massive data.

Aiming at the problems of the energy consumption of 802.15.4 network, a channel access mechanism for Media Access Control (MAC) layer based on multi-condition sleep mode was proposed. First, a Markov model of the mechanism was established. Then, the mathematical derivation based on the model of the steady-state probability of the main state, related parameters were given out. Furthermore, the analysis of the node average energy consumption in superframe was carried out. At last, the influence of the protocol parameters such as arrival rate of packets, number of back, superframe order and mininum of backoff exponent to the steady-state probability of the main state, the average energy consumption and the survival time of node was researched. The experimental results show that, compared with 802.15.4 network without node sleep state, the node energy consumption is reduced by 84.4% or so. And compared with the methods of some conditions, node energy consumption is reduced by 62.8% on average; the average survival time of network is increased by 70%. The model describes the proposed channel access mechanism very well, and the reasonable settings for parameters can improve the performance of node energy consumption. It also provides reference for the energy optimization in the practical application of Wireless Sensor Network (WSN).

In order to improve the transformation efficiency of tile-pyramid image, a 15-parameter projection transformation method was established by quartic polynomial based on the view model of digital earth. The influencing factors for selecting the size of tile image were discussed theoretically, and an optimization method to determine the size and depth of tile-pyramid was given. To test this algorithm, a basic digital earth environment BDE2 was constructed by adopting JOGL. The analysis and experimental results show that tile-pyramid in 10m pixel accuracy constructed by this algorithm only has 10 layers and less than 5×10-5 average error; meanwhile, the proposed algrithm has low complexity, close stitching, high definition and low distortion, and can effectively avoid stitch cracks and characteristics distortion after the image is transformed.

In the pointer forwarding schemes of Hierarchical Mobile IPv6 (HMIPv6), the influence of the distance between the Access Routers (ARs) on the handover performance has not been taken into consideration. To solve this problem, the optimization of intra-MAP (Mobile Anchor Point) handover in HMIPv6 based on pointer forwarding (OPF-HMIPv6) was proposed. The OPF-HMIPv6 compared the distance between ARs with the distance between AR and MAP firstly and gave priority to registering to MAP, rather than built a pointer chain immediately by registering to AR. The simulation results have shown that OPF-HMIPv6 can decrease the registration cost by 39% compared to HMIPv6 when the distance between AR and MAP is greater than the distance between ARs, which proves that the optimization reduces the overhead caused by the binding update and improves the efficiency of the intra-MAP handover.

Concerning the Internet news tracking, the study put forward a dynamic model for event tracking with reference to the time information. The dynamic model introduced the time factor into the traditional vector model to get the time similarity of the same characteristic words between the document and the event,and then applied the time similarity to calculate the similarity of the document and the event.If a document was related to the event,the new characteristic words in the document would be added to the event term set,and the weight and time information of characteristic words in the event term set should be re-adjusted. The experiment was evaluated by Detection Error Tradeoff (DET), and the results show that the dynamic model for event tracking improves the system performance effectively, and its minimum normalized cost of tracking loss is reduced by about 9%.

Traditional Collaborative Filter (CF) often suffers from the shortage of historic information. A transfer-Bagging algorithm based on hybrid classifiers was proposed for question recommendation. The main idea was that the recommendation and prediction problem were cast into the framework of transfer learning, then the users' demand for recommend questions were treated as target domain, while similar users who had applicable historic information were employed as auxiliary domain to help training target classifiers. The experimental results on both question recommendation platform and popular open datasets show that the accuracy of the proposed algorithm is 10%-20% higher than CF, and 5%-10% higher than single Bagging algorithm. The method solves cold start-up and sparse data problem in question recommendation field, and can be generalized into production recommendation on E-commerce platform.

Congestion window of TCP Westwood (TCPW) is based on the increase of linear mode at the congestion avoidance phase in high-speed networks. Therefore, it cannot rapidly obtain or maintain the high throughput. During the slow-start stage, the congestion window of TCPW is based on exponential growth mode, which will cause the datagram increases too fast and prompt the probability of congestion. For the above defects, TCPW was improved from two aspects, and the new algorithm was called NLTCPW. During the slow-start stage, send window of NLTCPW got 10 packets faster than TCPW. After that, the increment speed of send window was decelerated. A simple nonlinear mode was used to increase the congestion widow at the congestion stage. The performance analysis of mathematical model and simulation results show that NLTCPW algorithm has better throughput performance, lower packet loss rate and better fairness, and it is friendly and stable in high-speed networks.

The macro handover has caused high packet loss and long handover latency in Hierarchical Mobile IPv6 (HMIPv6) protocol. To solve these problems, this paper proposed a protocol named Tunnel-based Fast Macro-Handover (TBFMH), which introduced the mechanism of tunnel, acquired care-of addresses on the grounds of handover information, conducted duplication address detection in advance and completed local binding update while building the tunnels. The simulation results show that TBFMH can decrease the handover latency by 50% at least and reduce the packet loss rate compared to HMIPv6, which effectively improves the performance in the macro handover.

In order to improve the splice site recognition accuracy of Fuzzy Support Vector Machine (FSVM), a new method for computing the membership degree of sample was proposed. The initial membership was defined as the distance ratio of the sample to the two cluster centers of positive and negative samples, K-Nearest Neighbor (KNN) was adopted to compute the tightness of the samples, and the multiplication of the tightness and the initial membership degree was used as the ultimate membership. It will not only improve the membership degree of support vector, but also reduce the membership degree of noise sample. This method was applied to recognize the splice site, and the experimental results show that the recognition accuracy of constitutive 5′ and 3′ splice site reaches 94.65% and 88.97% respectively. Compared with the classical support vector machine，the recognition accuracy of constitutive 3′ splice site increases by 7.94%.

To meet the need of instruction detection system used in complex natural environment, such as coastal mudflats, an improved barrier coverage model, a multi-line barrier coverage scheduling protocol named k-MLBCSP, a coverage layout algorithm and a coverage adjustment algorithm were proposed. The k-MLBCSP protocol divided the network lifetime into three phases. In the initialization phase, the coverage layout algorithm guaranteed reasonable network settings. In the adjustment phase, the coverage adjustment algorithm provided an effective way for the sink and alive senosrs to further negotiate coverage layout strategies. The theoretical analysis and simulations show that compared with LBCP and RIS, k-MLBCSP increases the sensor network's coverage probability and lifetime. Furthermore, k-MLBCSP reduces the time complexity and the network load.

Independent Component Analysis (ICA) is a method for blind source separation based on higher-order statistics. It is hard to deal with the signal in the environment of Gaussian noise, because the higher-order cumulant of Gaussian signal is zero. A noisy image separation algorithm based on Curvelet threshold de-noising processing and FastICA was proposed. The results of simulation in Gaussian noise show that it can solve the problem of performance deterioration of ICA algorithms while processing noisy mixtures. Curvelet transform used in noisy images separation can improve the quality of Signal-to Noise Ratio (SNR) and the performance of separation compared with ICA that has been de-noised by wavelet.