A low-than character feature embedding called radical embedding is proposed, and applied on a long-short term memory (LSTM) model for sentence segmentation of pre-modern Chinese texts. The dataset includes over 150 classical Chinese books from 3 different dynasties and contains different literary styles. LSTM-conditional random fields (LSTM-CRF) model is a state-of-the-art method for the sequence labeling problem. This model adds a component of radical embedding, which leads to improved performances. Experimental results based on the aforementioned Chinese books demonstrate better accuracy than earlier methods on sentence segmentation, especial in Tang’s epitaph texts (achieving an F1-score of 81.34%).

Breast cancer is the most common cancer among women worldwide. Ultrasound is widely used as a harmless test for early breast cancer screening. The ultrasound network (USNet) model is presented. It is an improved object detection model specifically for breast nodule detection on ultrasound images. USNet improved the backbone network, optimized the generation of feature maps, and adjusted the loss function. Finally, USNet trained with real clinical data. The evaluation results show that the trained model has strong nodule detection ability. The mean average precision (mAP) value can reach 0.734 9. The nodule detection rate is 95.11%, and the in situ cancer detection rate is 79.65%. At the same time, detection speed can reach 27.3 frame per second (FPS), and the
video data can be processed in real time.

A new semi-serial fusion method of multiple feature based on learning using privileged information (LUPI) model was put forward. The exploitation of LUPI paradigm permits the improvement of the learning accuracy and its stability, by additional information and computations using optimization methods. The execution time is also reduced, by sparsity and dimension of testing feature. The essence of improvements obtained using multiple features types for the emotion recognition (speech expression recognition), is particularly applicable when there is only one modality but still need to improve the recognition. The results show that the LUPI in unimodal case is effective when the size of the feature is considerable. In comparison to other methods using one type of features or combining them in a concatenated way, this new method outperforms others in recognition accuracy, execution reduction, and stability.

In gesture recognition, static gestures, dynamic gestures and trajectory gestures are collectively known as multi-modal gestures. To solve the existing problem in different recognition methods for different modal gestures, a unified recognition algorithm is proposed. The angle change data of the finger joints and the movement of the centroid of the hand were acquired respectively by data glove and Kinect. Through the preprocessing of the multi-source heterogeneous data, all hand gestures were considered as curves while solving hand shaking, and a uniform hand gesture recognition algorithm was established to calculate the Pearson correlation coefficient between hand gestures for gesture recognition. In this way, complex gesture recognition was transformed into the problem of a simple comparison of curves similarities. The main innovations: 1) Aiming at solving the problem of multi-modal gesture recognition, an unified recognition model and a new algorithm is proposed; 2) The Pearson correlation coefficient for the first time to construct the gesture similarity operator is improved. By testing 50 kinds of gestures, the experimental results showed that the method presented could cope with intricate gesture interaction with the 97.7% recognition rate.

In case of machine learning, the problem of class imbalance is always troubling, i. e. one class of the samples has a larger magnitude than the other classes. This problem brings a preference of the classifier to the majority class, which leads to worse performance of the classifier on the minority class. We proposed an improved boosting tree (BT) algorithm for learning imbalanced data, called cost BT. In each iteration of the cost BT, only the weights of the misclassified minority class samples are increased. Meanwhile, the error rate in the weight formula of the base classifier is replaced by 1 minus F-measure. In this study, the performance of the cost BT algorithm is compared with other known methods on 9 public data sets. The compared methods include the decision tree and
random forest algorithm, and both of them were combined with the sampling techniques such as synthetic minority oversampling technique (SMOTE), Borderline-SMOTE, adaptive synthetic sampling approach (ADASYN) and one sided selection. The cost BT algorithm performed better than the other compared methods in F-measure, G-mean and area under curve (AUC). In 6 of the 9 data sets, the cost BT algorithm has a superior performance to the other published methods. It can promote the prediction performance of the base classifiers by increasing the proportion of the minority class in the whole samples with only increasing the weights of the misclassified minority class samples in each iteration of the BT. In addition, computing the weights of the base classifiers with F-measure is helpful to the ensemble decisions.

Video description aims to generate descriptive natural language for videos. Inspired from the deep neural network (DNN) used in the machine translation, the video description (VD) task applies the convolutional neural network (CNN) to extracting video features and the long short-term memory (LSTM) to generating descriptions. However, some models generate incorrect words and syntax. The reason may because that the previous models only apply LSTM to generate sentences, which learn insufficient linguistic information. In order to solve this problem, an end-to-end DNN model incorporated subject, verb and object (SVO) supervision is proposed. Experimental results on a publicly available dataset, i. e. Youtube2Text, indicate that our model gets a 58.4% consensus-based image
description evaluation (CIDEr) value. It outperforms the mean pool and video description with first feed (VD-FF) models, demonstrating the effectiveness of SVO supervision.

A joint hybrid beamforming and power splitting (JHBPS) design problem for simultaneous wireless information and power transfer (SWIPT) in millimeter-wave (mmWave) system is studied. The considered scenario is a multi-antenna base station (BS) transfers information and energy simultaneously to multiple single-antenna receivers. BS adopts hybrid digital and analog beamforming architecture to reduce hardware costs. Receivers separate acquired signals with power splitters either for information decoding (ID) or energy harvesting (EH). The aim is minimizing total transmission power by joint design of hybrid beamforming and PS under ID and EH requirements. It is difficult to obtain the optimal hybrid beamformer directly since the analog beamformer and digital beamformer are
multiplied. Therefore, a two-stage algorithm is proposed to solve the problem. In the first stage, the optimal beamformer and PS ratios are obtained by solving the joint transmission beamforming and PS design problem. In the second stage, the optimal beamformer is approximated with the product of analog beamformer and digital beamformer. The superiority of proposed algorithm over the existing algorithms is demonstrated through simulations. Moreover, the effectiveness of approximation algorithm is testified.

The cellular heterogeneous network (HetNet) with ultra dense small cells is called ultra cellular HetNet. The energy efficiency for this network is very important for future green wireless communications. The data rates and power consumptions for three parts (i.e., macro cells, small cells, and mixed backhaul links) in ultra cellular HetNet are jointly formulated to model downlink energy efficiency considering the active base stations (BSs) and inactive BSs. Then, in order to decrease the downlink co-channel interference, the interference price functions are also jointly set up for the three parts in ultra cellular HetNet. Next, energy efficiency optimization iterative algorithm scheme using the fractional programming and Lagrangian multiplier with constraints for density of ultra dense small cells and fraction of mixed backhaul links is presented with interference pricing. The convergence and computation complexity are also proved in this scheme. The numerical simulations finally demonstrate convergence behavior of the proposed algorithm. By comparison, some conclusion can be drawn. Maximizing energy efficiency of system is lower as the density of small cell is high. The effect on maximizing energy efficiency with interference price outperforms that without interference price. And the energy efficiency increases as the fraction of mixed backhaul links is higher because of more power consumption in the microwave backhaul links.

To progressively provide the competitive rate-distortion performance for aerial imagery, a quantized block compressive sensing (QBCS) framework is presented, which incorporates two measurement-side control parameters: measurement subrate (S) and quantization depth (D). By learning how different parameter
combinations may affect the quality-bitrate characteristics of aerial images, two parameter allocation models are derived between a bitrate budget and its appropriate parameters. Based on the corresponding allocation models, a model-guided image coding method is proposed to pre-determine the appropriate (S, D) combination for acquiring an aerial image via QBCS. The data-driven experimental results show that the proposed method can achieve near-optimal quality-bitrate performance under the QBCS framework.

To reduce fetching cost from a remote source, it is natural to cache information near the users who may access the information later. However, with development of 5G ultra-dense cellular networks andmobile edge computing (MEC), a reasonable selection among edge servers for content delivery becomes a problem when the mobile edge obtaining sufficient replica servers. In order to minimize the total cost accounting for both caching and fetching process, we study the optimal resource allocation for the content replica servers’deployment. We decompose the total cost as the superposition of cost in several coverages. Particularly, we consider the criterion for determining the coverage of a replica server and formulate the coverage as a tradeoff between caching cost and fetching cost. According to the criterion, a coverage isolation (CI) algorithm is proposed to solve the deployment problem. The numerical results show that the proposed CI algorithm can reduce the cost and obtain a higher tolerance for different centrality indices.