基于金字塔块匹配的双目场景流估计

针对现有双目场景流计算方法在大位移、运动遮挡及光照变化等复杂场景下场景流估计的准确性与鲁棒性问题,提出一种基于金字塔块匹配的双目场景流计算方法 .首先对双目图像序列进行超像素分割和视差估计,得到图像初始分割结果和视差信息,然后建立基于金字塔块匹配的运动模型并采用Ransac随机一致性算法拟合刚性运动模型和最小化重投影算法估计对象运动参数.最后,本文将金字塔块匹配结果作为约束项,联合对象运动参数和超像素平面参数构建基于金字塔块匹配的双目场景流估计能量函数模型,通过最小化能量函数得到最终场景流.实验分别采用KITTI2015(Karlsruhe Institute of Technology and Toyota technological Institute 2015)和MPI-Sintel(Max-Planck Institute and Sintel)数据集测试图像对本文方法和具有代表性场景流算法进行综合对比分析,结果表明本文方法相对于其他对比方法有效提高大位移、运动遮挡以及光照变化情况下场景流估计精度和鲁棒性.     

FRFCM聚类与深度优化的RGBD场景流计算

针对现有RGBD场景流计算模型在复杂场景、非刚性运动和运动遮挡等情况下易产生场景过度平滑和运动边缘模糊的问题,提出一种基于FRFCM（Fast and Robust Fuzzy C-Means）聚类与深度优化的RGBD场景流计算方法.首先以图像序列连续帧间光流信息为基准,利用FRFCM聚类算法对输入图像进行初始分割,然后根据深度图像的运动边缘信息优化初始分割结果,提取高置信度的运动分层信息.最后设计基于图像分割的RGBD场景流能量函数,采用金字塔变形策略计算精确的场景流结果.分别采用Middlebury和MPI-Sintel数据库所提供的测试图像集对本文方法和现有的RGBD场景流算法进行综合对比分析,实验结果表明本文方法相对于其他方法具有更好的场景流估计精度和鲁棒性,有效改善了场景过度平滑和运动边缘模糊问题.     

联合遮挡约束与残差补偿的特征金字塔光流计算方法

针对现有深度学习光流计算模型在运动遮挡和大位移等场景下光流计算的准确性与鲁棒性问题，本文提出一种联合遮挡约束与残差补偿的特征金字塔光流计算方法 .首先，构造基于遮挡掩模的光流约束模块，通过预测遮挡掩模特征图抑制变形特征的边缘伪影，克服运动遮挡区域的图像边缘模糊问题.然后，采用特征图变形策略构建基于特征变形的光流残差补偿模块，利用该模块学习到的残差光流细化原始光流场，改善大位移运动区域的光流计算效果.最后，采用特征金字塔框架构建联合遮挡约束与残差补偿的光流计算网络模型，提升大位移和运动遮挡场景下的光流计算精度.分别采用MPI-Sintel (Max-Planck Institute and Sintel)和KITTI (Karlsruhe Institute of Technology and Toyota Technological Institute)数据集对本文方法和代表性传统光流计算方法、深度学习光流计算方法进行综合对比分析，实验结果表明本文方法相对于其他方法能够有效提升大位移和运动遮挡场景下的光流计算精度与鲁棒性.     

基于多阈值归一化分割的模糊图像边缘分割算法

模糊图像边缘的像素特征较为复杂,一般需要采用多个阈值作为分隔约束条件的方法来进行图像边缘分割,但是该方法存在诸如多阈值无法形成统一标准、边缘提取过程需要多次校对,以及效率较低等缺点.提出一种基于多阈值归一化分割的模糊图像边缘分割算法,通过设计超像素网格对模糊图像边缘特征的像素进行匹配,分析模糊图像的反调张量信息,并根据不同张量信息对多阈值进行归一化,以及采用灰度窗口相关系数匹配方法,将获得的多阈值归一化结果分别覆盖图中的单一目标对象,以实现模糊图像的边缘分割.实验表明,利用该算法进行模糊图像边缘分割能较好地获取图像的边缘细节特征,使得边缘具有更好的连线段连通性和宽度一致性.     

一种边缘保持立体匹配方法

针对边缘处前景和背景视差易混淆问题,提出一种边缘保持立体匹配方法.在代价匹配阶段,采用级联Census变换增强代价的抗噪特性.在代价聚集阶段,引入SLIC超像素分割信息进行快速边缘保持代价聚集.在视差后处理阶段,通过导向十字滤波器进一步优化边缘视差.实验结果表明,文中提出的立体匹配方法在Middlebury测试集以及实际场景获得高质量视差效果,并在边缘处的视差较以往非局部立体匹配方法有所提升.实验还发现在点云上采样时,引入本文所提的导向十字滤波器,可以解决点云在边缘处的过渡.     

一种视频对象分割新方法

提出一种半自动的视频对象分割新方法,通过对象跟踪分割视频序列的后继帧。这种方法首先采用基于块匹配和最大边缘强度的运动估值和补偿方法进行了对象轮廓定位,接着采用模板匹配以特定对象知识检测对象像素。为使轮廓定位更可靠,在块匹配的运动估值中使用了彩色信息。而模板匹配则使分割结果精确化,避免误差传递,并且在出现遮挡时只要对象颜色在整个序列中的一直保持相似性,就能够正确检测出对象。实验结果证明这种方法能够分割复杂场景中的任意对象。     

一种新的多尺度深度学习图像语义理解方法研究

如何在深度学习中融合 图像的多尺度信息,是基于深度学习的视觉算法需要解决的一个关键问题。本文提出一种基 于多尺度交替 迭代训练的深度学习方法,并应用于图像的语义理解。算法采用卷积神经网络(CNN)从原始 图像中提取稠密性特征 来编码以每个像素为中心的矩形区域,将多个尺度图像交替迭代训练,能够捕获不同尺度下 的纹理、颜色和 边缘等重要信息。在深度学习提取特征分类结果的基础上,提出了一种结合超像素分割的方 法,统计超像 素块的主导类别,来校正分类错误的像素类别,同时描绘出目标区域边界轮廓,完成最终的 语义理解。在Stanford Background Dataset 8类数据集上验证了本文方法的有效性,准确 率达到77.4%。     

基于深度卷积神经网络的红外场景理解算法

采用深度学习的方法实现红外图像场景语义理解。首先,建立含有4类别前景目标和1个类别背景的用于语义分割研究的红外图像数据集。其次,以深度卷积神经网络为基础,结合条件随机场后处理优化模型,搭建端到端的红外语义分割算法框架并进行训练。最后,在可见光和红外测试集上对算法框架的输出结果进行评估分析。实验结果表明,采用深度学习的方法对红外图像进行语义分割能实现图像的像素级分类,并获得较高的预测精度。从而可以获得红外图像中景物的形状、种类、位置分布等信息,实现红外场景的语义理解。     

语义驱动的颜色恒常决策算法

不同颜色恒常性算法适用于不同场景下的图像,算法融合是扩展颜色恒常性算法适用范围常用的方法之一,而现有融合性算法在算法选择依据上忽略了语义信息在图像纹理特征描述中的作用,导致光源估计时的精度不高。针对该问题,提出一种语义驱动的颜色恒常决策算法。首先,利用PSPNet(Pyramid Scene Parsing Network)模型对经过一阶灰度边缘算法(1st Gray Edge)偏色预处理后的目标图像进行场景语义分割,并计算场景中各个语义类别的占比;其次,根据语义类别及占比在已训练的决策集合中寻找相似的参考图像,并使用欧氏距离计算两者的语义相似度;最后,将语义相似度与基于多维欧氏空间确定的阈值进行判别,根据判别结果选择合适算法为目标图像实行偏色校正。在Color Checker和NUS-8 camera两种数据集中的实验结果表明,所提算法光源估计角度误差较单一算法均大幅度下降,且较同类型融合性算法分别下降14.02%和8.17%,提高了光源估计的鲁棒性和准确度。     

基于改进的SLIC的岩心颗粒图像边缘分割算法

在岩心颗粒图像进行目标提取的过程中,由于颗粒颜色丰富,类别和大小不一,且存在边界模糊等情况,导致颗粒分割很困难.针对以上问题,本文提出一种基于改进的简单线性迭代聚类(SLIC)算法,首先对图像进行预处理,增强目标区域同时模糊背景部分,消除孤立的噪声点且保护边缘信息;其次,结合LBP纹理特征对图像进行超像素分割;最后,结合区域之间的颜色特征进行超像素合并.实验表明,与现有的其它算法相比,该算法能准确地分割颗粒的边界,更有效地提取目标颗粒,极大地降低了后续对提取目标进行分析的复杂度.     

Study on traffic scene semantic segmentation method based on convolutional neural network

In order to improve the semantic segmentation accuracy of traffic scene,a segmentation method was proposed based on RGB-D image and convolutional neural network.Firstly,on the basis of semi-global stereo matching algorithm,the disparity map was obtained,and the sample library was established by fusing the disparity map D and RGB image into the four-channel RGB-D image.Then,with two different structures,the networks were trained by using two different learning rate adjustment strategy respectively.Finally,the traffic scene semantic segmentation test was carried out with RGB-D image as the input,and the results were compared with the segmentation method based on RGB image.The experimental results show that the proposed traffic scene segmentation algorithm based on RGB-D image can achieve higher semantic segmentation accuracy than that based on RGB image.     

Superpixels extracted via region fusion with boundary constraint

In this paper, we present an accurate superpixel algorithm by region fusion with boundary constraint (RFBC). Superpixels with regular shape and high boundary adherence can be generated in weak boundary and complex texture regions through our algorithm. RFBC includes two steps which are initial segmentation and region fusion respectively. In initial segmentation, broken Canny edges are connected through edge closing algorithm. Subsequently, the closed Canny edges and SLIC superpixel edges are combined together to form the incipient superpixels. In region fusion, gray Gaussian distribution and adjacent relation are used as priori to compute the degree of similarity across incipient superpixels in GBP algorithm. For concreteness, the information of similarity is propagated between regions and the most similar regions are fused, which are accomplished alternatingly to preserve accurate boundaries. Extensive experiments on the Berkeley segmentation benchmark show that the proposed algorithm outperforms the most state-of-the-art algorithms.     

基于光流与Delaunay三角网格的图像序列运动遮挡检测

针对图像序列运动遮挡检测的准确性与鲁棒性问题,提出一种基于光流与Delaunay三角网格的图像序列运动遮挡检测方法.首先构造基于非局部约束的TV-L¹光流估计模型;然后根据图像Delaunay三角网格划分与光流估计结果对图像序列帧间对应像素点和局部三角形进行运动遮挡判断并检测遮挡区域;最后采用MPI Sintel和Middlebury数据库提供的测试图像集对本文方法与SMOD、GOSF等代表性方法进行对比测试.实验结果表明,本文方法相对于SMOD和GOSF方法在十组测试图像集的平均漏检率和误检率分别降低15.21%与30.57%,说明本文方法针对非刚性运动、复杂场景、弱纹理、光照阴影以及大位移等类型图像序列均具有较高的检测精度和较好的鲁棒性.     

改进U-Net网络的接地网图像超像素分割北大核心CSCD

研究基于改进U-Net网络的接地网图像超像素分割方法,提升红外图像超像素分割效果。通过主成分分析法降维处理接地网腐蚀红外图像;利用Turbopixel超像素分割法分割降维后的红外图像,获取数个超像素区域;在全卷积U-Net网络内添加可变形卷积与重构上采样卷积,并利用反向传播算法,优化网络参数,建立改进的全卷积U-Net网络结构;在改进的全卷积U-Net网络内分割获取的数个超像素区域,输出红外图像超像素自动分割结果。实验证明:该方法可有效降维处理接地网腐蚀红外图像,实现红外图像超像素分割,分割后的红外图像边界清晰;在不同分辨率时,该方法的Dice相似性系数较高、Hausdorff距离较低,具备较高的红外图像超像素分割精度。     

SCTMS: Superpixel based color topographic map segmentation method

Different from natural image, topographic map is a complex manually generated image which has amount of interlaced lines and area features. Because of the frequent intersection and the overlap between geographic elements, the misalignment in scanner and other disturbances like inappropriate preserving, false color, mixed color and color aliasing problems occur in the raster color maps. These problems could cause serious challenges in segmentation process. In this work, we present a color topographic map segmentation method based on superpixel to overcome these problems. Firstly, the finest partition is obtained based on double color-opponent boundary detection method and watershed approach. Then, a strict region merging method is introduced to prevent mis-merging while superpixels generated. This merging method could make the superpixel partition accurately adherent the boundary between different geographic elements. Finally, luminosity, color and texture information are combinative applied to classify the superpixel into different layers based on support vector machine. The experimental results show that the proposed method outperforms other state-of-art topographic map segmentation approaches.     

相互结构引导滤波TV-L1变分光流估计

由于光流场既包含物体的运动信息,又包含场景的三维结构信息,因此光流计算技术是计算机视觉和机器视觉领域研究的重要任务之一.针对现有光流计算方法在图像边缘保护方面存在过度平滑问题,提出一种基于相互结构引导滤波的TV-L1（Total Variational with L1 norm,TV-L1）变分光流估计方法.通过提取置信度较高的图像相互结构区域,构造基于相互结构引导滤波的全局目标函数,并采用金字塔分层细化与交替迭代方案结合的策略进行优化,该方法可以较好的保护图像边缘信息.最后采用标准测试图像集对本文方法与现有代表性变分方法LDOF（Large Displacement Optical Flow,LDOF）,CLG-TV（Combined Local-Global Total Variation,CLG-TV）,Classic++,NNF（Nearest Neighbor Fields,NNF）以及深度学习方法FlowNet2.0进行对比,实验结果表明本文方法具有较高的光流估计精度与鲁棒性,尤其对图像边缘保护具有显著的效果,并且在运动目标检测,机器人避障等方面具有一定应用前景.     

Depth segmentation in real-world scenes based on U–V disparity analysis

Depth segmentation has the challenge of separating the objects from their supporting surfaces in a noisy environment. To address the issue, a novel segmentation scheme based on disparity analysis is proposed. First, we transform a depth scene into the corresponding U-V disparity map. Then, we conduct a region-based detection method to divide the object region into several targets in the processed U-disparity map. Thirdly, the horizontal plane regions may be mapped as slant lines in the V-disparity map, the Random Sample Consensus (RANSAC) algorithm is improved to fit such multiple lines. Moreover, noise regions are reduced by image processing strategies during the above processes. We respectively evaluate our approach on both real-world scenes and public data sets to verify the flexibility and generalization. Sufficient experimental results indicate that the algorithm can efficiently segment and label a full-view scene into a group of valid regions as well as removing surrounding noise regions.     

Depth estimation of supervised monocular images based on semantic segmentation

In recent years, the research method of depth estimation of target images using Convolutional Neural Networks (CNN) has been widely recognized in the fields of artificial intelligence, scene understanding and three-dimensional (3D) reconstruction. The fusion of semantic segmentation information and depth estimation will further improve the quality of acquired depth images. However, how to deeply combine image semantic information with image depth information and use image edge information more accurately to improve the accuracy of depth image is still an urgent problem to be solved. For this purpose, we propose a novel depth estimation model based on semantic segmentation to estimate the depth of monocular images in this paper. Firstly, a shared parameter model of semantic segmentation information and depth estimation information is built, and the semantic segmentation information is used to guide depth acquisition in an auxiliary way. Then, through the multi-scale feature fusion module, the feature information contained in the neural network on different layers is fused, and the local feature information and global feature information are effectively used to generate high-resolution feature maps, so as to achieve the goal of improving the quality of depth image by optimizing the semantic segmentation model. The experimental results show that the model can fully extract and combine the image feature information, which improves the quality of monocular depth vision estimation. Compared with other advanced models, our model has certain advantages.