Sequential alignment attention model for scene text recognition

Scene text recognition has been a hot research topic in computer vision due to its various applications. The state-of-the-art solutions usually depend on the attention-based encoder-decoder framework that learns the mapping between input images and output sequences in a purely data-driven way. Unfortunately, there often exists severe misalignment between feature areas and text labels in real-world scenarios. To address this problem, this paper proposes a sequential alignment attention model to enhance the alignment between input images and output character sequences. In this model, an attention gated recurrent unit (AGRU) is first devised to distinguish the text and background regions, and further extract the localized features focusing on sequential text regions. Furthermore, CTC guided decoding strategy is integrated into the popular attention-based decoder, which not only helps to boost the convergence of the training but also enhances the well-aligned sequence recognition. Extensive experiments on various benchmarks, including the IIIT5k, SVT, and ICDAR datasets, show that our method substantially outperforms the state-of-the-art methods.     

ATSP:时态网络可视化的自适应时间片划分方法

现有的时态网络可视化方法大多采用等量时间片来可视化网络的演变，不利于时态模式的快速挖掘和发现。为此，根据时态网络固有的特征提出自适应时间片划分方法（Adaptive Time Slice Partition method，ATSP）。在时态网络的两种表示方式（基于事件的表示方式和基于快照的表示方式）的基础上，构建了ATSP的基础模型，同时提出了一种改进模型用来描述事件间隔时间服从长尾分布的时态网络。为了实现时间片的不等量划分，针对探索任务的不同提出了基于时态模式的ATSP规则和基于中心节点的ATSP规则，并提出了实现算法--层次划分算法（Hierarchical Partition algorithm，HP）和增量划分算法（Incremental Partition algorithm，IP）。实验结果表明，ATSP方法比传统的时间片划分方法更能准确地表示网络的时态特征，且该方法应用于可视化时，能有效归纳并展示网络的特征，明显提高了视觉分析的效率。     

时态神经网络模型及其在股票分类预测上应用

对具有时间属性的数据进行数据挖掘称为时态数据挖掘，用以发现数据在时间上的知识，当数据变化不规律时，如股票交易数据,就很难发现有价值的规律与规则。而神经网络具有并行、容错、可以硬件实现以及自我学习的优点,可作为股票分类预测应用的一种方法。通过将股票数据与时态型相结合，将股票数据转换成时态型股票数据，提出时态神经网络模型的分类方法，对收集的若干上市公司十年内的股票数据进行分析，构建了时态股票数据神经网络分类器对股票进行分类预测。经过实验验证，相比改进前的神经网络和支持向量机方法，该分类器具有更高的分类准确率。结果证明，这种时态数据神经网络模型对于多只股票的分类预测是非常有效的，可以很好地运用到股票市场的分类预测中。     

On the capacity of the state‐dependent interference relay channel

This paper considers the state‐dependent interference relay channel (SIRC) in which one of the two users may operate as a secondary user and the relay has a noncausal access to the signals from both users. For discrete memoryless SIRC, we first establish the achievable rate region by carefully merging Han‐Kobayashi rate splitting encoding technique, superposition encoding, and Gelfand‐Pinsker encoding technique. Then, based on the achievable rate region that we derive, the capacity of the SIRC is established in many different scenarios including (a) the weak interference regime, (b) the strong interference regime, and (c) the very strong interference regime. This means that our capacity results contain all available known results in the literature. Next, the achievable rate region and the associated capacity results are also evaluated in the case of additive Gaussian noise. Additionally, many numerical examples are investigated to show the value of our theoretical derivations.     

Cloud resource prediction model based on triple exponential smoothing method and temporal convolutional network

The container cloud represented by Docker and Kubernetes has the advantages of less additional resource overhead and shorter start-up and destruction time.However there are still resource management issues such as over-supply and under-supply.In order to allow the Kubernetes cluster to respond “in advance” to the resource usage of the applications deployed on it,and then to schedule and allocate resources in a timely,accurate and dynamic manner based on the predicted value,a cloud resource prediction model based on triple exponential smoothing method and temporal convolutional network was proposed,based on historical data to predict future demand for resources.To find the optimal combination of parameters,the parameters were optimized using TPOT thought.Experiments on the CPU and memory of the Google dataset show that the model has better prediction performance than other models.     

Differences between oculomotor and perceptual artifacts for temporally limited head mounted displays

We used perceptual and oculomotor measures to understand the negative impacts of low (phantom array) and high (motion blur) duty cycles with a high‐speed, AR‐likehead‐mounted display prototype. We observed large intersubject variability for the detection of phantom array artifacts but a highly consistent and systematic effect on saccadic eye movement targeting during low duty cycle presentations. This adverse effect on saccade endpoints was also related to an increased error rate in a perceptual discrimination task, showing a direct effect of display duty cycle on the perceptual quality. For high duty cycles, the probability of detecting motion blur increased during head movements, and this effect was elevated at lower refresh rates. We did not find an impact of the temporal display characteristics on compensatory eye movements during head motion (e.g., VOR). Together, our results allow us to quantify the tradeoff of different negative spatiotemporal impacts of user movements and make subsequent recommendations for optimized temporal HMD parameters.     

Optimal weighed Holoentropy for video compression: an advanced HEVC

ABSTRACT

In recent times, the applications of multimedia are rising in a greedy mode and hence the amount of video transactions is also increasing exponentially. This has shouted great demands on effective models on video encoding and also for reducing the transmission channel congestion. This research work introduces a managing technique termed weighted encoding for High-Efficiency Video Coding (HEVC). HEVC, also termed as MPEG-H Part 2 and H.265 is a video compression standard that is widely utilized AVC (H.264 or MPEG-4 Part 10). When compared to AVC, HEVC grants double the ratio of data compression at a similar level of quality of the video or considerably enhanced video quality at a similar bit rate. This work intends to optimize the weight that adopted in HEVC for encoding. For this, this paper proposes a new Iterative based propagation update in the water wave Optimization Algorithm (IPU-WWO), which is the improved form of Water wave Optimization (WWO). The performance of proposed IPU-WWO is compared over other conventional methods like Artificial Bee Colony (ABC), Firefly (FF), Particle Swarm Optimization (PSO) and Genetic Algorithm (GA) with respect to Peak Signal to Noise Ratio (PSNR). By doing the encoding process, it minimizes the video size with perceptually better quality video or PSNR.     

通风量对菇渣好氧发酵温度时空特性的影响

采用不同通风量进行菇渣与猪粪高温共发酵,研究通风对发酵过程中堆体温度的时空变化的影响。结果表明,通风使堆体温度呈现出明显的梯度效应,不同处理各发酵时期的温度分布均由剖面中部向四周不断降低。高温区域的位置和面积随发酵时间呈动态变化规律,升温期温度在发酵罐中上部积聚,高温分布区类似椭圆形,高温期椭圆面积逐渐向堆体中部扩散和扩大,随后高温面积不断递减并向中下部迁移,降温期堆体各区域温度趋于一致。由等值线的疏密程度可知,温度的空间变异性大体呈现出高温期>第2次高温期>升温期>降温期的规律。结合试验各阶段的温度变化并考虑节能降耗,得出发酵升温期和高温期分别以0.05,0.067m3/(d·m3)进行通风,二次高温期与降温期以0.033 m3/(d·m3)进行通风条件最佳。     

基于深度卷积长短时神经网络的视频帧预测

针对视频帧预测中难以准确预测空间结构信息细节的问题，通过对卷积长短时记忆（LSTM）神经网络的改进，提出了一种深度卷积长短时神经网络的方法。首先，将输入序列图像输入到两个不同通道的深度卷积LSTM网络组成的编码网络中，由编码网络学习输入序列图像的位置信息变化特征和空间结构信息变化特征；然后，将学习到的变化特征输入到与编码网络通道数对应的解码网络中，由解码网络输出预测的下一张图；最后，将这张图输入回解码网络中，预测接下来的一张图，循环预先设定的次后输出全部的预测图。与卷积LSTM神经网络相比，在Moving-MNIST数据集上的实验中，相同训练步数下所提方法不仅保留了位置信息预测准确的特点，而且空间结构信息细节表征能力更强。同时，将卷积门控循环单元（GRU）神经网络的卷积层加深后，该方法在空间结构信息细节表征上也取得了提升，检验了该方法思想的通用性。     

Temporal and spatial evolution measurement of laser-induced breakdown spectroscopy on hydrogen retention in tantalum

Fuel retention measurement on plasma-facing components is an active field of study in magnetic confinement nuclear fusion devices.The laser-induced breakdown spectroscopy(LIBS)diagnostic method has been well demonstrated to detect the elemental distribution in PFCs.In this work,an upgraded co-axis LIBS system based on a linear fiber bundle collection system has been developed to measure the hydrogen(H) retention on a tantalum(Ta) sample under a vacuum condition.The spatial resolution measurement of the different positions of the LIBS plasma can be achieved simultaneously with varying delay times.The temporal and spatial evolution results of LIBS plasma emission show that the H plasma observably expands from the delay times of 0-200 ns.The diameter of Ta plasma is about 6 mm which is much less than the size of H plasma after 200 ns.The difference in the temporal and spatial evolution behaviors between H plasma and Ta plasma is due to the great difference in the atomic mass of H and Ta.The depth profile result shows that H retention mainly exists on the surface of the sample.The temporal and spatial evolution behaviors of the electron excited temperature are consistent with that of the Ta emission.The result will further improve the understanding of the evolution of the dynamics of LIBS plasma and optimize the current collection system of in situ LIBS in fusion devices.