To better apply deep convolutional neural networks for expression recognition in UAV-enabled B5G/6G networks, we propose a deep network expression recognition method based on transfer learning and fine-tuning on facial expression datasets. Initially, we train our model on a large-scale facial attribute dataset and subsequently fine-tune it on a facial expression dataset. This strategy effectively lowers the high costs and dependence associated with annotating facial expression datasets, while enhancing the accuracy and training speed of our model. This method is particularly suited for UAVs equipped with on-board cameras and image processing capabilities, enabling real-time expression recognition for various applications such as crowd monitoring, search and rescue, and human-UAV interaction. Compared to traditional methods that train exclusively on facial expression datasets, our method significantly reduces the number of training iterations on facial expression datasets and significantly improves the generalization ability of the model, especially for UAV applications. We use a large-scale facial attribute dataset, which is more closely related to the facial expression recognition task, as our source dataset, forming a contrast with methods that typically use a facial recognition dataset as the transfer learning source dataset. The experimental results on the CK + dataset, integrated with UAV-enabled B5G/6G networks, show that our method achieves a facial expression recognition accuracy of 97.6%, a significant improvement over the 97.3% accuracy rate of methods that only train on facial expression datasets, with less training time as well.

Benefiting from deep learning, the accuracy of face expression recognition tasks based on convolutional neural networks has been greatly improved. However, the traditional SoftMax activation function lacks the ability to discriminate between classes. To solve this problem, the industry has proposed several activation functions based on softmax, such as A-softmax, LMCL, etc. We investigate the geometric significance of the weights from a fully connected layer and consider the weights as the class centers. By extracting the feature vector of several samples and extending the corresponding means to the weights, the model can develop the ability to recognize custom classes without training, while maintaining the accuracy of the original classification. On the expression task, the original seven-category classification is validated to obtain 97.10% accuracy on the CK+ dataset and 88% accuracy on the custom dataset.

With the application of 5G technology in the field of education, the construction of smart campus has set off a wave of digital transformation. At the same time, the traditional smart campus is also facing the exponential growth of the number of Internet of Things devices, servers, and application terminals, which makes it difficult to achieve flat management. In view of the current difficulties in the construction of smart campus, this paper proposes smart campus architecture based on blockchain technology. Unlike the traditional smart campus architecture, this paper combines the characteristics of decentralization, high confidentiality, and data sharing of block chain with the Internet of Things technology, which greatly reduces the demand for data storage and physical network equipment. The new smart campus architecture plans the application of smart education based on blockchain, and provides a new solution model and research ideas.

The use of the internet of things (IoT) is steadily increasing in a wide range of applications. Integration of IoT, computer vision, and artificial intelligence can improve people's daily life in various domains such as smart homes, smart cities, and smart industries. There are a large number of face recognition and face attribute recognition scenarios in reality, and the industry commonly decomposes these tasks, with three models responsible for handling face detection, face recognition, and face attribute recognition. The multi-model approach requires a lot of computational resources for context switching, while training one model with one dataset is not only complex, but also leads to overfitting of the multi-model approach. The authors propose a single-model multi-task approach, which can complete all tasks using only one model, and thus obtains a great improvement in inference speed, especially in scenes with high face density. After an experimental comparison, our approach saves a maximum of 96% of inference time, 49.5% of memory usage, and 59.7% of CPU time, avoids frequent context switching, and simplifies the training steps while improving the generalization performance of the model.

Recognizing defects in X-ray images plays an important role in the detection of internal defects in titanium alloy castings. However, the existing manual defects recognition methods have common drawbacks such as unstable artificial recognition, misrecognition, huge workload, and low efficiency of recognition. To make up for the shortcomings, an ameliorated deep dense convolutional neural network (BX-Net) was presented to accurately recognize casting defects in X-ray images and effectively extract highly discriminative features of different categories. DenseNet121 was used as the backbone of BX-Net and the feature extracted by DenseNet121 was fully shared by the two inputs of a bilinear pooling layer. Transfer learning was applied to reduce the demand for data and hyperparameters’ tuning. The backbone of BX-Net was firstly trained on the ImageNet and then all layers of BX-Net was fine-tuned on nine-hundred X-ray images of TiAl aero casting components. Other six deep convolutional neural networks (DenseNet121, EfficientNetB4, EfficientNetB7, ResNet50, VGG16 and Xception) were also trained to be compared with the presented BX-Net. Two Support Vector Machines were trained on the LBP features data set and HOG features data set of nine-hundred X-ray images of TiAl aero casting components respectively Experiments comparing BX-Net with other six deep learning models and two machine learning models on one hundred X-ray images (test set) of TiAl aero casting components were carried out. The comparison results show that BX-Net has the least parameters except for the Densenet121. The recall and accuracy of BX-Net were 99% and 99% respectively. In addition, the comparison results also show that BX-Net was the only model that learned discriminative feature representation of the casting X-ray image data set. The BX-Net proposed in this paper is expected to overcome the shortcomings of manual defects recognition.

本研究提出适合高职在线教育的"教学模式-质量监管-安全保障"方案,以现有资源为中心,同步直播讲授教学、在线SPOC教学和SPOC+直播混合教学3种在线教学模式;构建大数据综合诊断及分析平台,针对web日志、资源使用、课程建设、教学运行等多维数据进行大数据智能分析,实现有效督导,确保教学质量;采取统一身份认证、VPN网络、运维审计、风险控制等技术保障网络安全.最后根据高职院校教育教学特点,对本方案提出反思,并针对后疫情时代在线教育方法、平台建设提出一些建议,为全国高职院校开展在线教育提供参考.

Intensity inhomogeneities often cause considerable difficulties in the quantitative analysis of images, especially for magnetic resonance (MR) images. It is hard to segment such images by using traditional level set based upon segmentation models. In order to come over these difficulties, in view of the experience of the predecessors, we proposed a new region-based active contour model. This model uses adaptive region force and the object's boundary information to segment images we greatly interested in. The region-based force has local information which can handle intensity inhomogeneity, the time-consumption is also optimistic. The GAC model plays a role which can help to get through the weak boundary. Experimental results for synthetic and real images demonstrate the superior performance of our model.

Learning and planning in partially observable Markov decision processes (POMDPs) is computationally intractable in real-time system. In order to address this problem, this paper proposes a belief policy reuse (BPR) method to avoid repeated computation. Firstly, the policy reuse evaluation mechanism based on belief Kullback-Leibler divergence is presented as a similarity metric between beliefs in the belief-policy library. If the current belief is similar to any of the past ones, the policy in the belief-policy library is reused. Otherwise, BPR exploits Monte-Carlo particle method to explore a new policy, and stores the new policy with belief in the belief-policy library, so it can be reused in the future. The experimental results show that the proposed approach is an effective way for improving the learning efficiency in large-scale partially observable Markov decision processes.

Bayesian reinforcement learning provides an elegant solution to the optimal tradeoff between exploration and exploitation of the uncertainty in learning. Unfortunately, the size of the learning parameters grows exponentially with the problem horizon. In this paper, we propose a novel Monte Carlo tree search for Bayesian reinforcement learning approach using a compact factored representation, to solve the Bayesian reinforcement learning problem online. At first, we exploit a factored representation to describe the states and transition function to reduce the size of learning parameters. Then, we can formulate model-based Bayesian reinforcement learning as a partially observable Markov decision processes. At last, we we apply the partially observable Monte-Carlo planning method as an online solver for Bayesian reinforcement learning. The experimental results show that the proposed approach is an effective way for improving the learning efficiency in large-scale Bayesian reinforcement learning.

A good skin detector that is capable of capturing skin tones under different conditions is important for human-machine interaction applications. In a general situation, skin detectors, such as skin probability maps or Gaussian mixture models, achieve acceptable skin segmentation results. However, the false positive rate increases significantly when the skin tones are in shadow or when skin-like background objects are under similar illumination. In this paper, we propose a novel skin feature learning algorithm based on stacked autoencoders, which are deep neural networks. To overcome the problems encountered in skin segmentation that are caused by different ethnicities and varying illumination conditions, the stacked autoencoders are utilized to learn more discriminative representations of the skin area in both the RGB color space and the HSV color space. Unlike traditional machine learning methods, instead of predicting each pixel individually, our algorithm utilizes blocks to learn the representations and detect the skin areas. The algorithm exploits the learning ability of deep neural networks to learn high-level representations of skin tones. Experiments on test images show that the proposed algorithm achieves acceptable results on several publicly available data sets. To reduce the difficulty of detecting skin pixels in these data sets, the ground truths of these data sets are commonly focused on foreground skin area detection. Our skin detector is also able to detect background areas, as shown in our experiments.

Monte-Carlo tree search (MCTS) combines the generality of stochastic simulation and the accuracy of tree search, which has attracted the great attention of scholars. However, the MCTS search requires a sufficient number of iterations to converge to a good solution, which is more difficult to optimize. In order to solve this problem, this paper presents a point-based incremental pruning (PIP) for Monte-Carlo tree search. Instead of reasoning about the whole policy trees space when pruning the cross-sums of the value functions during policy construction, our algorithm uses boundary belief points to perform exact pruning, and exploits intermediate points to perform approximate pruning by generating policy trees, then uses real-time belief states to get the optimal policy in policy execution. The theoretical analysis and empirical results indicate that PIP can speed up the tree search efficiently.

The sensor scheduling for energy-efficient target tracking with high performance in wireless sensor networks (WSNs) is a dilemma problem. By analyzing the intrinsic relationship between tracking performance and energy consumption, we cast the scheduling problem of WSN as the optimal policy problem of partially observable Markov decision process (POMDP), and propose a dynamic cluster members scheduling (DCMS) algorithm to solve the tradeoff between tracking performance and energy consumption. First, we exploit an election method, based on the optimal mixed weights of the signal strength and the residual energy of node, to choose the cluster head node. Then, we seem each cluster members an agent, and model the scheduling problem of cluster members by POMDP. At last, a point-based online value iteration algorithm is presented to solve the DCMS to generate the collaboration strategy of sensor cluster members dynamically. The simulation results show that the proposed approach can improve the accuracy of target tracking, decrease the energy consumption of sensor nodes, and prolong the lifetime of sensor networks.

In multiple-input multiple-output (MIMO) communication systems, complex matrix inversion is a very computationally demanding operation. Especially when the number of antennas increases, i.e., in a massive MIMO system, the complexity of matrix inversion becomes very high. Motivated by this observation, a new low complexity complex matrix inversion method, called SDF-SGR (Square root and Division Free Squared Givens' Rotations) based algorithm, is designed for MIMO channels. Square root operation is avoided in the whole algorithm, and division operation is replaced by shift operation during the Givens' rotations phase. Besides, since the scale factor z, involved in the traditional SGR, has little influence on the whole algorithm, it is removed in this SDF-SGR algorithm. Considering a 4×4 complex matrix inversion, the SDF-SGR based algorithm could reduce the multiplication operations by 13.07% compared with the traditional SGR algorithm, and the division operation is reduced by almost 52.94%.

不管是建设整体数字校园系统，还是部分系统，都要做好整体规划工作。

Partially Observable Markov Decision Processes (POMDPs) provide a natural and principled framework for sequential decision-making under uncertainty. However, the large-scale POMDPs suffer from the exponential growth of policy trees space. In order to solve this problem, we present a novel Point-based Incremental Pruning (PIP) algorithm based on the piecewise linearity and convexity of value functions. Instead of reasoning about the whole policy trees space when pruning the cross-sums of the value functions during policy construction, our algorithm uses boundary belief points to perform exact pruning, and exploits intermediate points to perform approximate pruning by generating policy trees, then use real-time belief states to get the optimal policy in policy execution. The empirical results indicate that PIP can handle large-scale POMDPs domains efficiently. 1553-9105/Copyright © 2014 Binary Information Press.

Taking the issue of information island in the field of security surveillance into consideration,this paper proposes a big data-based distributed cloud computing model,consisting of domain node,sub-domain nodes and terminal nodes.The terminal nodes are responsible for data collection,the domain node and sub-domain nodes take charge of data computation and data storage.Then we present an architecture of security surveillance big data based on cloud computing.Characterized by low communication bandwidth,high elasticity of extension,and easy implementation,the proposed computing model integrates the advantages of big data and cloud computing,solves the problem of information island in the field of security surveillance.