A Transformer-Based Capsule Network for 3D Part–Whole Relationship Learning

Learning the relationship between the part and whole of an object, such as humans recognizing objects, is a challenging task. In this paper, we specifically design a novel neural network to explore the local-to-global cognition of 3D models and the aggregation of structural contextual features in 3D space, inspired by the recent success of Transformer in natural language processing (NLP) and impressive strides in image analysis tasks such as image classification and object detection. We build a 3D shape Transformer based on local shape representation, which provides relation learning between local patches on 3D mesh models. Similar to token (word) states in NLP, we propose local shape tokens to encode local geometric information. On this basis, we design a shape-Transformer-based capsule routing algorithm. By applying an iterative capsule routing algorithm, local shape information can be further aggregated into high-level capsules containing deeper contextual information so as to realize the cognition from the local to the whole. We performed classification tasks on the deformable 3D object data sets SHREC10 and SHREC15 and the large data set ModelNet40, and obtained profound results, which shows that our model has excellent performance in complex 3D model recognition and big data feature learning.     

基于支持向量机增量学习和LPBoost的人体目标再识别算法

摄像机间目标关联是无重叠视域多摄像机目标持续跟踪的关键.提出了一种只利用人体目标外观,完全不依赖于空时关系的人体目标再识别算法,利用识别结果直接进行跨摄像机间人体目标关联,而不依赖于目标的捕获时间和路径限制.对跟踪视频前景图像序列提取互补性视觉单词树直方图和全局颜色直方图二种特征,采用支持向量机增量学习在线训练二种特征的人体外观辨别模型,再利用多类线性规划增强算法对二种特征的支持向量机模型进行在线自适应融合.实验结果表明,本文算法具有较强的在线学习能力,能增量式表达人体目标辨别性外观模型,特征融合后的模型区别性更强,有效地降低多方面条件变化的影响,获得了高识别率,且能够实现快速实时实现,相对于现有方法有了明显提升.     

Fast aircraft detection using cascaded discriminative model in photoelectric sensing system

Aircraft detection is a fundamental problem in computer vision. As a vision-based system, the photoelectric sensing system (in airport) needs to capture the aircrafts quickly and accurately by the optical camera. Although many existing detection models reach to favorable accuracy, they are time consuming in training and testing, which is not suitable for this system. In practice, as a core part of vision-based system, detection module always occupies a lot of time in image processing and target matching. To reduce the (detection) time cost without losing detection accuracy, we designed a cascade discriminative model which includes two stages: coarse pre-detection stage and fine detection stage. In the traditional object detection models, generally, an object feature template was employed to search for all positions and levels in image pyramid with sliding window fashion. However, in our detection model, only a small number of candidate regions were pre-detected to reduce the searching space at the first stage. At the second stage, an assembled method (which includes partitioned bag-of-words method and random forest) was adopted for accelerating the feature quantization and formation. Then, the possible regions including object were decided by a non-linear SVM classifier. We evaluated our model on two benchmark databases (Caltech 101 and PASCAL 2007) and our own database (images were obtained from the optical camera), and it yields high performance. Compared with other state-of-the-art methods, our model outperforms them not only in detection speed, but also in detection accuracy.     

Infrared dim moving target tracking via sparsity-based discriminative classifier and convolutional network

Infrared dim and small target tracking is a great challenging task. The main challenge for target tracking is to account for appearance change of an object, which submerges in the cluttered background. An efficient appearance model that exploits both the global template and local representation over infrared image sequences is constructed for dim moving target tracking. A Sparsity-based Discriminative Classifier (SDC) and a Convolutional Network-based Generative Model (CNGM) are combined with a prior model. In the SDC model, a sparse representation-based algorithm is adopted to calculate the confidence value that assigns more weights to target templates than negative background templates. In the CNGM model, simple cell feature maps are obtained by calculating the convolution between target templates and fixed filters, which are extracted from the target region at the first frame. These maps measure similarities between each filter and local intensity patterns across the target template, therefore encoding its local structural information. Then, all the maps form a representation, preserving the inner geometric layout of a candidate template. Furthermore, the fixed target template set is processed via an efficient prior model. The same operation is applied to candidate templates in the CNGM model. The online update scheme not only accounts for appearance variations but also alleviates the migration problem. At last, collaborative confidence values of particles are utilized to generate particles' importance weights. Experiments on various infrared sequences have validated the tracking capability of the presented algorithm. Experimental results show that this algorithm runs in real-time and provides a higher accuracy than state of the art algorithms.     

Saliency detection based on diffusion maps

Visual saliency has recently attracted lots of research interest in the computer vision community. In this paper, we propose a novel computational model for bottom-up saliency detection based on manifold learning. A typical graph-based manifold learning algorithm, namely the diffusion maps, is adopted for establishing our saliency model. In the proposed method, firstly, a graph is constructed using low-level image features. Then, the diffusion maps algorithm is performed to learn the diffusion distances at different time, which are utilized to derive the saliency measure. Compared to existing saliency models, our method has the advantage of being able to capture the intrinsic nonlinear structures in the original feature space. Experimental results on publicly available data demonstrate that our method outperforms the state-of-the-art saliency models, both qualitatively and quantitatively.     

Saliency detection via two-directional 2DPCA analysis of image patches

Salient object detection is an important and challenging problem in computer vision. In this paper, we present a model of salient region detection based on the fusion of contrast and distribution, computed by two-directional 2DPCA analysis of image patches under the combination of RGB space, LAB space and YCbCr space. First, non-overlap patches of three layers from the image are obtained in the three color spaces respectively and stacked for the combination of the three sapces in a single layer. For every layer, two-directional, two-dimensional PCA are utilized to realize automatic selection of effective features, then based on the high contrast and compact character of salient object, contrast values and distribution values of image patches are fused to get the saliency map. Finally, three saliency maps for three layers are combined to detect salient object. The experimental results on a publicly available database show that the proposed algorithm performs well and are in line with the human eye observation results.     

基于稀疏活动轮廓模型的感兴趣目标检测定位算法

传统的基于形状信息目标定位的算法,对目标观测角度发生形变情况下的定位存在不少困难,针对该问题,提出了一种基于稀疏活动轮廓模型的感兴趣目标(OOI)检测算法.首先通过共同勾画算法学习到感兴趣目标的稀疏活动轮廓模型,它能够清晰地定义感兴趣目标模式;同时构成该模型的Gabor轮廓基元可以通过扰动进行局部的调整以适配图像,在一...     

基于SOM和RBF网络的软测量方法研究

针对氢粉碎过程中钕铁硼粉碎状态不可知,为有效预测合金的反应状态,提出了一种基于自组织特征映射(SOM)神经网络和径向基函数(RBF)神经网络结合构建的网络模型。在该模型中,SOM神经网络作为聚类网络,采用无教师学习算法对输入样本进行自组织分类,并将分类中心及其对应的权值向量传递给RBF神经网络,作为径向基函数的中心;RBF神经网络作为基础网络,采用高斯函数作为径向基函数实现从输入到隐含层的非线性映射,输出层则采用有教师学习算法训练网络的权值,从而实现输入层到输出层的线性映射。并以钕铁硼氢粉碎过程合金中氢含量为检测对象,运用上述方法在MATLAB平台上建立了合金中氢含量预测模型,并完成了仿真验证。     

Robustness quantification method for network intrusion detection models

Deep learning techniques have been successfully applied to network intrusion detection tasks, but as in the case of autonomous driving and face recognition, the reliability of the system itself has become a pressing issue. Robustness is a key attribute to determine whether a deep learning system is secure and reliable, and we also choose to explore the security of intrusion detection models from a new perspective of robustness quantification. In this paper, we focus on the intrusion detection model based on long and short-term memory, and use a fine-grained linear approximation method to derive a more accurate robustness bound on the nonlinear activation function with tighter linear constraints. We can use this bound to quantitatively measure the robustness of the detection model and determine whether the model is susceptible to the influence of adversarial samples. In our experiments, we test networks with various structures on the MNIST dataset, and the results show that our proposed method can effectively deduce the robustness bounds of output elements, and has good scalability and applicability.     

自动发音错误检测中基于F1值最大化的声学模型训练方法

摘要为了提高计算机辅助语言学习中自动发音错误检测系统的性能,提出一种声学模型的区分性训练方法。该方法将经过正确度标注的非母语语音数据库上的发音错误检测的F₁值的最大化作为模型参数的训练准则。采用Sigmoid 函数对F₁值函数进行平滑构造目标函数,并利用构造弱意义辅助函数的方法以及扩展Baum-Welch 形式的参数更新公式进行优化。提出在模型参数更新与音素门限同时优化的策略保证目标函数增长的单调性。发音错误检测实验表明该方法能够有效地增大训练和测试数据检错的F₁值。同时训练数据和测试数据上的精确度、召回率以及检测正确度都有明显改进。      

Transformation properties of the Lyapunov exponent of dynamical systems under diffeomorphisms

In this paper we present a geometric definition of the Lyapunov exponent on a differential manifold and investigate its transformation properties under changes of coordinates, or, more generally, under diffeomorphisms. The result is that not every diffeomorphism leaves the Lyapunov exponent invariant. A sufficient condition for invariance is the following: the tangent map of the diffeomorphism is bounded exponentially in the curve parameter for any curve in the manifold and any direction in the tangent bundle with basis restricted to this curve. At the end we show that for a free particle there are diffeomorphisms violating this condition, although they are even canonical maps.     

An object tracking method based on guided filter for night fusion image

Online object tracking is a challenging problem as it entails learning an effective model to account for appearance change caused by intrinsic and extrinsic factors. In this paper, we propose a novel online object tracking with guided image filter for accurate and robust night fusion image tracking. Firstly, frame difference is applied to produce the coarse target, which helps to generate observation models. Under the restriction of these models and local source image, guided filter generates sufficient and accurate foreground target. Then accurate boundaries of the target can be extracted from detection results. Finally timely updating for observation models help to avoid tracking shift. Both qualitative and quantitative evaluations on challenging image sequences demonstrate that the proposed tracking algorithm performs favorably against several state-of-art methods.     

层叠式“产生/判别”混合模型的语音情感识别

提出了层叠式“产生/判别”混合模型的语音情感识别方法。首先,提取63维语句级特征,运用Fisher从中选择12个最佳的语句级特征,建立小波神经网络(WNN)的层叠式产生式模型进行语音情感识别;然后提取69维帧级特征,采用SFS选择出待使用的8维特征,将高斯混合模型(GMM)进行多维概率输出,建立层叠式“产生/判别”混合模型进行语音情感识别。实验结果显示:(1)层叠式“产生/判别”混合模型较单独WNN、GMM、HMM (隐马尔可夫模型)、SVM (支持向量机)的识别率要高;(2)层叠式“产生/判决式”混合模型识别率较基于WNN的层叠产生式模型高;(3) M=13,D维GMM-MAP/SVM (MAP,最大后验概率)串联融合模型为最优的层叠式“产生/判别”混合模型,能获得最高85.1%的识别率。      

The valuation of equity warrants in a fractional Brownian environment

In this paper, we discuss the valuation of equity warrants in the geometric fractional Brownian environment based on the equilibrium condition. Using the conditional expectation we present a fractional pricing model for equity warrants and analyze the influence of the Hurst parameter. Then we propose an optimization procedure to obtain the valuation of equity warrants. Some numerical examples are given to demonstrate the pricing results by comparing different pricing models. Furthermore, we provide an empirical study to show how to apply our model in realistic contexts, and these comparative results of different pricing models show that the pricing model proposed in this paper matches the actual price quite well.     

Using deep neural networks to diagnose engine pre-ignition

Engine downsizing and boosting have been recognized as effective strategies for improving engine efficiency. However, operating the engines at high load promotes abnormal combustion events, such as pre-ignition and potential superknock. Currently the most effective method for detecting pre-ignition is by using in-cylinder pressure sensors that have high precision and sensitivity, but also high cost. Due to rapid advances in automotive technology such as autonomous driving, computer-aided designs and future connectivity, we propose to use a complimentary data-driven strategy for diagnosing abnormal combustion events. To this end, a data-driven diagnostics approach for pre-ignition detection with deep neural networks is proposed. The success of convolutional neural networks (CNNs) in object detection and recurrent neural networks (RNNs) in sequence forecasting inspired us to develop these models for pre-ignition detection. For a cost-effective strategy, we use data from less expensive sensors, such as lambda and low-resolution exhaust back pressure (EBP), instead of high resolution in-cylinder pressure measurements. The first deep learning model is combined with a commonly used dimensionality reduction tool–Principal Component Analysis (PCA). The second model eliminates this step and directly processes time-series data. Results indicate that the first model with reduced input dimensions, and correspondingly smaller size of the network, shows better performance in detecting pre-ignition cycles with an F1 score of 79%. Overall, the proposed deep learning approach is a promising alternative for abnormal combustion diagnostics using data from low resolution sensors.     

Dynamic 3-D shape measurement method based on FTP

In this paper, we propose a method for dynamic three-dimensional (3-D) shape measurement based on Fourier transform profilometry. A sequence of dynamic deformed fringe images can be grabbed by a CCD camera and saved on a disk rapidly. By a Fourier transform, filtering and inverse Fourier transform, a sequence of phase-maps can be obtained. By unwrapping these phase maps in 3-D phase space, we can obtain the shape of the dynamic object at different times. In this paper we also propose the algorithm of a phase difference between two deformed fringes, and the 3-D phase unwrapping method based on the phase difference algorithm. The computer simulation and experiment results show that this method can efficiently deal with the dynamic 3-D shape measurement.     

Protein transfer to membranes upon shape deformation

Red blood cells, milk fat droplets, or liposomes all have interfaces consisting of lipid membranes. These particles show significant shape deformations as a result of flow. Here we show that these shape deformations can induce adsorption of proteins to the membrane. Red blood cell deformability is an important factor in several diseases involving obstructions of the microcirculatory system, and deformation induced protein adsorption will alter the rigidity of their membranes. Deformation induced protein transfer will also affect adsorption of cells onto implant surfaces, and the performance of liposome based controlled release systems. Quantitative models describing this phenomenon in biomaterials do not exist. Using a simple quantitative model, we provide new insight in this phenomenon. We present data that show convincingly that for cells or droplets with diameters upwards of a few micrometers, shape deformations induce adsorption of proteins at their interface even at moderate flow rates.     

Estimating time to contact during pursuit eye movements: Comparison between geometric model prediction and human performance

While an object is approaching a particular location, we can make an estimate of the time when the object will arrive at that location. A geometric model predicts that the estimate of time-to-contact (TTC) is greatly improved by using the rate of change of visual direction of the object when the object is moving with a slow velocity toward a point of nearest approach at a distance far from the observer. It has been shown that pursuit eye movements provide the rate of change of visual direction of an approaching object. We conducted psychophysical experiments, and compared TTC estimates during pursuit eye movements to those during fixation. We found that the differences in TTC estimates between fixation and pursuit show a qualitatively similar pattern to the geometric model prediction. However, the results also show that the magnitudes of the TTC estimation errors are greater than the theoretical values from the geometric model, indicating that the human visual system has a perceptual bias in estimating TTC. These results suggest that the human visual system estimates TTC during pursuit eye movements in a different way from the geometric model, although the effect of these eye movements on TTC estimates in human performance is qualitatively consistent with the model prediction.