基于拼接缝自适应消除和全景图矫直的快速图像拼接算法

针对图像拼接存在色差过渡不均匀、图像倾斜扭曲及拼接效率低的现象，提出一种基于拼接缝自适应消除和全景图矫直的快速图像拼接算法。首先，用尺度不变特征变换（SIFT）提取图像指定区域特征点并用双向K近邻（KNN）算法进行图像配准，有效提高算法效率；其次，利用动态规划思想提出自适应公式找到最优拼接缝并用图像融合算法对其自适应消除，解决拼接缝色差过渡不均匀问题；最后，针对累积拼接误差形成全景图倾斜的现象，利用边缘检测算法提出自适应拟合四边形矫直模型，把原始全景图矫直为一个全新的全景图。所提算法与分块图像拼接和二叉树图像拼接算法相比，图像质量提升了5.84%~7.83%，拼接时间仅为原来的50%~70%。实验结果表明，该算法不仅通过自适应更新机制减少不同图像背景下拼接缝色差过渡不均匀的现象，从而提高了图像质量；而且提高了拼接效率，降低了全景图倾斜扭曲程度。     

A hybrid multiple feature construction approach for classification using Genetic Programming

The purpose of feature construction is to create new higher-level features from original ones. Genetic Programming (GP) was usually employed to perform feature construction tasks due to its flexible representation. Filter-based approach and wrapper-based approach are two commonly used feature construction approaches according to their different evaluation functions. In this paper, we propose a hybrid feature construction approach using genetic programming (Hybrid-GPFC) that combines filter’s fitness function and wrapper’s fitness function, and propose a multiple feature construction method that stores top excellent individuals during a single GP run. Experiments on ten datasets show that our proposed multiple feature construction method (Fcm) can achieve better (or equivalent) classification performance than the single feature construction method (Fcs), and our Hybrid-GPFC can obtain better classification performance than filter-based feature construction approaches (Filter-GPFC) and wrapper-based feature construction approaches (Wrapper-GPFC) in most cases. Further investigations on combinations of constructed features and original features show that constructed features augmented with original features do not improve the classification performance comparing with constructed features only. The comparisons with three state-of-art methods show that in majority of cases, our proposed hybrid multiple feature construction approach can achieve better classification performance.     

Facial expression recognition of intercepted video sequences based on feature point movement trend and feature block texture variation

Facial Expression Recognition (FER) is an important subject of human–computer interaction and has long been a research area of great interest. Accurate Facial Expression Sequence Interception (FESI) and discriminative expression feature extraction are two enormous challenges for the video-based FER. This paper proposes a framework of FER for the intercepted video sequences by using feature point movement trend and feature block texture variation. Firstly, the feature points are marked by Active Appearance Model (AAM) and the most representative 24 of them are selected. Secondly, facial expression sequence is intercepted from the face video by determining two key frames whose emotional intensities are minimum and maximum, respectively. Thirdly, the trend curve which represents the Euclidean distance variations between any two selected feature points is fitted, and the slopes of specific points on the trend curve are calculated. Finally, combining Slope Set which is composed by the calculated slopes with the proposed Feature Block Texture Difference (FBTD) which refers to the texture variation of facial patch, the final expressional feature are formed and inputted to One-dimensional Convolution Neural Network (1DCNN) for FER. Five experiments are conducted in this research, and three average FER rates 95.2%, 96.5%, and 97% for Beihang University (BHU) facial expression database, MMI facial expression database, and the combination of two databases, respectively, have shown the significant advantages of the proposed method over the existing ones.     

Probabilistic tree-based representation for solving minimum cost integer flow problems with nonlinear non-convex cost functions

The minimum cost flow problem (MCFP) is the most generic variation of the network flow problem which aims to transfer a commodity throughout the network to satisfy demands. The problem size (in terms of the number of nodes and arcs) and the shape of the cost function are the most critical factors when considering MCFPs. Existing mathematical programming techniques often assume the cost functions to be linear or convex. Unfortunately, the linearity and convexity assumptions are too restrictive for modelling many real-world scenarios. In addition, many real-world MCFPs are large-scale, with networks having a large number of nodes and arcs. In this paper, we propose a probabilistic tree-based genetic algorithm (PTbGA) for solving large-scale minimum cost integer flow problems with nonlinear non-convex cost functions. We first compare this probabilistic tree-based representation scheme with the priority-based representation scheme, which is the most commonly-used representation for solving MCFPs. We then compare the performance of PTbGA with that of the priority-based genetic algorithm (PrGA), and two state-of-the-art mathematical solvers on a set of MCFP instances. Our experimental results demonstrate the superiority and efficiency of PTbGA in dealing with large-sized MCFPs, as compared to the PrGA method and the mathematical solvers.     

基于高光谱图像的改进SIFT特征提取与匹配

针对尺度不变特征变换(SIFT)算法所提取图像特征点数量少、误匹率高的问题,提出了一种基于高光谱图像的改进SIFT算法。首先,依据传统SIFT算法中高斯金字塔的构造思想,结合在不同波段下的高光谱图像具有相同宏观特征的特点,首次用高光谱图像作为原始算法中经高斯变换产生的图像,使得检测到的具有实际意义的特征点数量大幅增加;其次,传统SIFT算法以及大量的改进方法都只通过目标象元邻域范围内的像素信息来构造特征描述符,而忽略了像素点的位置信息,文中将目标象元的位置信息纳入了特征描述符,在特征描述符的匹配阶段,在利用邻域范围内的像素信息进行粗匹配之后,利用特征描述符中的位置信息进行精细匹配。仿真实验结果表明在限定最优值与次优值之比的情况下,采用高光谱图像构造高斯金字塔的方式能显著增加特征点的提取数量,更多地挖掘出图像中的极值点;在特征描述符中加入目标象元的位置信息作为特征点匹配第二阶段的判断依据,正确匹配数量达到原方法的59倍以上,极大提升了算法的匹配性能。     

基于CapsNet的汉字字形表征模型

提出基于胶囊神经网络(CapsNet)的汉字字形表征模型,通过表征汉字字形中的部件实现汉字字形的表征.首先,对任一汉字字形生成所有部件类别的表征向量.然后,根据部件存在概率,利用基于欧氏距离的离群点检测,选取相应的部件表征向量.最后,由选出的部件表征向量组成该汉字的字形表征.实验表明,文中模型在仅经过部件字形训练的情况下,即可有效识别汉字部件,同时自动生成汉字字形的有效表征.     

基于特征选择和深度信念网络的文本情感分类算法

由于人类语言的复杂性，文本情感分类算法大多都存在因为冗余而造成的词汇量过大的问题。深度信念网络（DBN）通过学习输入语料中的有用信息以及它的几个隐藏层来解决这个问题。然而对于大型应用程序来说，DBN是一个耗时且计算代价昂贵的算法。针对这个问题，提出了一种半监督的情感分类算法，即基于特征选择和深度信念网络的文本情感分类算法（FSDBN）。首先使用特征选择方法（文档频率（DF）、信息增益（IG）、卡方统计（CHI）、互信息（MI））过滤掉一些不相关的特征从而使词汇表的复杂性降低；然后将特征选择的结果输入到DBN中，使得DBN的学习阶段更加高效。将所提算法应用到中文以及维吾尔语中，实验结果表明在酒店评论数据集上，FSDBN在准确率方面比DBN提高了1.6%，在训练时间上比DBN缩短一半。