提取全局语义信息的场景图生成算法

目的 场景图能够简洁且结构化地描述图像。现有场景图生成方法重点关注图像的视觉特征,忽视了数据集中丰富的语义信息。同时,受到数据集长尾分布的影响,大多数方法不能很好地对出现概率较小的三元组进行推理,而是趋于得到高频三元组。另外,现有大多数方法都采用相同的网络结构来推理目标和关系类别,不具有针对性。为了解决上述问题,本文提出一种提取全局语义信息的场景图生成算法。方法 网络由语义编码、特征编码、目标推断以及关系推理等4个模块组成。语义编码模块从图像区域描述中提取语义信息并计算全局统计知识,融合得到鲁棒的全局语义信息来辅助不常见三元组的推理。目标编码模块提取图像的视觉特征。目标推断和关系推理模块采用不同的特征融合方法,分别利用门控图神经网络和门控循环单元进行特征学习。在此基础上,在全局统计知识的辅助下进行目标类别和关系类别推理。最后利用解析器构造场景图,进而结构化地描述图像。结果 在公开的视觉基因组数据集上与其他10种方法进行比较,分别实现关系分类、场景图元素分类和场景图生成这3个任务,在限制和不限制每对目标只有一种关系的条件下,平均召回率分别达到了44.2%和55.3%。在可视化实验中,相比...     

Object Bank: An Object-Level Image Representation for High-Level Visual Recognition

It is a remarkable fact that images are related to objects constituting them. In this paper, we propose to represent images by using objects appearing in them. We introduce the novel concept of object bank (OB), a high-level image representation encoding object appearance and spatial location information in images. OB represents an image based on its response to a large number of pre-trained object detectors, or ‘object filters’, blind to the testing dataset and visual recognition task. Our OB representation demonstrates promising potential in high level image recognition tasks. It significantly outperforms traditional low level image representations in image classification on various benchmark image datasets by using simple, off-the-shelf classification algorithms such as linear SVM and logistic regression. In this paper, we analyze OB in detail, explaining our design choice of OB for achieving its best potential on different types of datasets. We demonstrate that object bank is a high level representation, from which we can easily discover semantic information of unknown images. We provide guidelines for effectively applying OB to high level image recognition tasks where it could be easily compressed for efficient computation in practice and is very robust to various classifiers.     

OPTIMOL: Automatic Online Picture Collection via Incremental Model Learning

The explosion of the Internet provides us with a tremendous resource of images shared online. It also confronts vision researchers the problem of finding effective methods to navigate the vast amount of visual information. Semantic image understanding plays a vital role towards solving this problem. One important task in image understanding is object recognition, in particular, generic object categorization. Critical to this problem are the issues of learning and dataset. Abundant data helps to train a robust recognition system, while a good object classifier can help to collect a large amount of images. This paper presents a novel object recognition algorithm that performs automatic dataset collecting and incremental model learning simultaneously. The goal of this work is to use the tremendous resources of the web to learn robust object category models for detecting and searching for objects in real-world cluttered scenes. Humans contiguously update the knowledge of objects when new examples are observed. Our framework emulates this human learning process by iteratively accumulating model knowledge and image examples. We adapt a non-parametric latent topic model and propose an incremental learning framework. Our algorithm is capable of automatically collecting much larger object category datasets for 22 randomly selected classes from the Caltech 101 dataset. Furthermore, our system offers not only more images in each object category but also a robust object category model and meaningful image annotation. Our experiments show that OPTIMOL is capable of collecting image datasets that are superior to the well known manually collected object datasets Caltech 101 and LabelMe.     

Adaptive image segmentation for region-based object retrieval using generalized Hough transform

Finding an object inside a target image by querying multimedia data is desirable, but remains a challenge. The effectiveness of region-based representation for content-based image retrieval is extensively studied in the literature. One common weakness of region-based approaches is that perform detection using low level visual features within the region and the homogeneous image regions have little correspondence to the semantic objects. Thus, the retrieval results are often far from satisfactory. In addition, the performance is significantly affected by consistency in the segmented regions of the target object from the query and database images. Instead of solving these problems independently, this paper proposes region-based object retrieval using the generalized Hough transform (GHT) and adaptive image segmentation. The proposed approach has two phases. First, a learning phase identifies and stores stable parameters for segmenting each database image. In the retrieval phase, the adaptive image segmentation process is also performed to segment a query image into regions for retrieving visual objects inside database images through the GHT with a modified voting scheme to locate the target visual object under a certain affine transformation. The learned parameters make the segmentation results of query and database images more stable and consistent. Computer simulation results show that the proposed method gives good performance in terms of retrieval accuracy, robustness, and execution speed.     

In the wild image retrieval and clustering for 3D cultural heritage landmarks reconstruction

One of the main characteristics of Internet era is the free and online availability of extremely large collections of images located on distributed and heterogeneous platforms over the web. The proliferation of millions of shared photographs spurred the emergence of new image retrieval techniques based not only on images’ visual information, but on geo-location tags and camera exif data. These huge visual collections provide a unique opportunity for cultural heritage documentation and 3D reconstruction. The main difficulty, however, is that the internet image datasets are unstructured containing many outliers. For this reason, in this paper a new content-based image filtering is proposed to discard image outliers that either confuse or significantly delay the followed e-documentation tools, such as 3D reconstruction of a cultural heritage object. The presented approach exploits and fuses two unsupervised clustering techniques: DBSCAN and spectral clustering. DBSCAN algorithm is used to remove outliers from the initially retrieved dataset and spectral clustering discriminate the noise free image dataset into different categories each representing characteristic geometric views of cultural heritage objects. To discard the image outliers, we consider images as points onto a multi-dimensional manifold and the multi-dimensional scaling algorithm is adopted to relate the space of the image distances with the space of Gram matrices through which we are able to compute the image coordinates. Finally, structure from motion is utilized for 3D reconstruction of cultural heritage landmarks. Evaluation on a dataset of about 31,000 cultural heritage images being retrieved from internet collections with many outliers indicate the robustness and cost effectiveness of the proposed method towards a reliable and just-in-time 3D reconstruction than existing state-of-the-art techniques.     

Segmentation-based multi-class semantic object detection

In this paper we study the problem of the detection of semantic objects from known categories in images. Unlike existing techniques which operate at the pixel or at a patch level for recognition, we propose to rely on the categorization of image segments. Recent work has highlighted that image segments provide a sound support for visual object class recognition. In this work, we use image segments as primitives to extract robust features and train detection models for a predefined set of categories. Several segmentation algorithms are benchmarked and their performances for segment recognition are compared. We then propose two methods for enhancing the segments classification, one based on the fusion of the classification results obtained with the different segmentations, the other one based on the optimization of the global labelling by correcting local ambiguities between neighbor segments. We use as a benchmark the Microsoft MSRC-21 image database and show that our method competes with the current state-of-the-art.     

Graph-based representation for similarity retrieval of symbolic images

Image retrieval from an image database by the image objects and their spatial relationships has emerged as an important research subject in these decades. To retrieve images similar to a given query image, retrieval methods must assess the similarity degree between a database image and the query image by the extracted features with acceptable efficiency and effectiveness. This paper proposes a graph-based model SRG (spatial relation graph) to represent the semantic information of the contained objects and their spatial relationships in an image with no file annotation. In an SRG graph, the image objects are symbolized by the predefined class names as vertices and the spatial relations between object pairs are represented as arcs. The proposed model assesses the similarity degree between two images by calculating the maximum common subgraph of two corresponding SRG’s through intersection, which has quadratic time complexity owing to the characteristics of SRG. Its efficiency remains quadratic regardless of the duplication rate of the object symbols. The extended model SRG_T is also proposed, with the same time complexity, for the applications that need to consider the topological relations among objects. A synthetic symbolic image database and an existing image dataset are used in the conducted experiments to verify the performance of the proposed models. The experimental results show that the proposed models have compatible retrieval quality with remarkable efficiency improvements compared with three well-known methods LCS_Clique, SIM_R, and 2D Be-string, where LCS_Clique utilizes the number of objects in the maximum common subimage as its similarity function, SIM_R uses accumulation-based similarity function of similar object pairs, and 2D Be-string calculates the similarity of 2D patterns by the linear combination of two 1D similarities.     

Semantic association enhancement transformer with relative position for image captioning

Transformer-based architectures have shown encouraging results in image captioning. They usually utilize self-attention based methods to establish the semantic association between objects in an image for predicting caption. However, when appearance features between the candidate object and query object show weak dependence, the self-attention based methods are hard to capture the semantic association between them. In this paper, a Semantic Association Enhancement Transformer model is proposed to address the above challenge. First, an Appearance-Geometry Multi-Head Attention is introduced to model a visual relationship by integrating the geometry features and appearance features of the objects. The visual relationship characterizes the semantic association and relative position among the objects. Secondly, a Visual Relationship Improving module is presented to weigh the importance of appearance feature and geometry feature of query object to the modeled visual relationship. Then, the visual relationship among different objects is adaptively improved according to the constructed importance, especially the objects with weak dependence on appearance features, thereby enhancing their semantic association. Extensive experiments on MS COCO dataset demonstrate that the proposed method outperforms the state-of-the-art methods.

     

基于细粒度视觉特征和知识图谱的视觉故事生成算法

视觉故事生成是图像内容描述衍生的跨模态学习任务,在图文游记自动生成、启蒙教育等领域有较好的应用研究意义。目前主流方法存在对图像细粒度特征描述薄弱、故事文本的图文相关性低、语言不丰富等问题。为此,该文提出了基于细粒度视觉特征和知识图谱的视觉故事生成算法。该算法针对如何对图像内容进行充分挖掘和扩展表示,在视觉和高层语义方面,分别设计实现了图像细粒度视觉特征生成器和图像语义概念词集合生成器两个重要模块。在这两个模块中,细粒度视觉信息通过含有实体关系的场景图结构进行图卷积学习,高层语义信息综合外部知识图谱与相邻图像的语义关联进行扩充丰富,最终实现对图像序列内容较为全面细致的表示。该文算法在目前视觉故事生成领域规模最大的VIST数据集上与主流先进的算法进行了测试。实验结果表明,该文所提算法生成的故事文本,在图文相关性、故事逻辑性、文字多样性等方面,在Distinct-N和TTR等客观指标上均取得较大领先优势,具有良好的应用前景。     

大规模图像集中的代表性图像选取

针对传统图像检索系统通过关键字搜索图像时缺乏语义主题多样性的问题,提出了一种基于互近邻一致性和近邻传播的代表性图像选取算法,为每个查询选取与其相关的不同语义主题的图像集合. 该算法利用互近邻一致性调整图像间的相似度,再进行近邻传播（AP）聚类将图像集分为若干簇,最后通过簇排序选取代表性图像簇并从中选取中心图像为代表性图像. 实验表明,本文方法的性能超过基于K-means的方法和基于Greedy K-means的方法,所选图像能直观有效地概括源图像集的内容,并且在语义上多样化.     

融合视觉关系检测的电力场景自动危险预警

目的 借助深度学习强大的识别与检测能力,辅助人工进行电力场景下的危险描述与作业预警是一种较为经济和高效的电力安全监管手段。然而,目前主流的以目标检测技术为基础的预警系统只能给出部分危险目标的信息,忽视了电力设备的单目危险关系和成对对象间潜在的二元危险关系。不同于以往的方法,为了拓展危险预警模块的识别能力与功能范畴,本文提出了一种在电力场景下基于视觉关系检测的自动危险预警描述生成方法。方法 对给定的待检测图像,通过目标检测模块得到图中对象的类别名称和限界框位置;分别对图像进行语义特征、视觉特征和空间位置特征的抽取,将融合后的总特征送入关系检测模块,输出单个对象的一元关系和成对对象间的关系三元组;根据检测出的对象类别和关系信息,进行危险预测并给出警示描述。结果 本文自主搜集了多场景下的电力生产作业图像并进行标注,同时进行大量消融实验。实验显示,结合了语义特征、空间特征和视觉特征的关系检测器在前5召回率Recall@5和前10召回率Recall@10上的精度分别达到86.80%和93.93%,比仅使用视觉特征的关系检测器的性能提高约15%。结论 本文提出的融合多模态特征输入的视觉关系检测网络能够较好地给出谓词关系的最佳匹配,并减少不合理的关系预测,且具有一定零样本学习（zero-shot learning）能力。相关可视化结果表明,整体系统能够较好地完成电力场景下的危险预警描述任务。     

一种适合弱标签数据集的图像语义标注方法

真实环境下数据集中广泛存在着标签噪声问题,数据集的弱标签性已严重阻碍了图像语义标注的实用化进程.针对弱标签数据集中的标签不准确、不完整和语义分布失衡现象,提出了一种适用于弱标签数据集的图像语义标注方法.首先,在视觉内容与标签语义的一致性约束、标签相关性约束和语义稀疏性约束下,通过直推式学习填充样本标签,构建样本的近似语义平衡邻域.鉴于邻域中存在噪声干扰,通过多标签语义嵌入的邻域最大边际学习获得距离测度和图像语义的一致性,使得近邻处于同一语义子空间.然后,以近邻为局部坐标基,通过邻域非负稀疏编码获得目标图像和近邻的部分相关性,并构建局部语义一致邻域.以邻域内的语义近邻为指导并结合语境相关信息,进行迭代式降噪与标签预测.实验结果表明了方法的有效性.     

深度卷积神经网络图像语义分割研究进展

在计算机视觉领域中,语义分割是场景解析和行为识别的关键任务,基于深度卷积神经网络的图像语义分割方法已经取得突破性进展。语义分割的任务是对图像中的每一个像素分配所属的类别标签,属于像素级的图像理解。目标检测仅定位目标的边界框,而语义分割需要分割出图像中的目标。本文首先分析和描述了语义分割领域存在的困难和挑战,介绍了语义分割算法性能评价的常用数据集和客观评测指标。然后,归纳和总结了现阶段主流的基于深度卷积神经网络的图像语义分割方法的国内外研究现状,依据网络训练是否需要像素级的标注图像,将现有方法分为基于监督学习的语义分割和基于弱监督学习的语义分割两类,详细阐述并分析这两类方法各自的优势和不足。本文在PASCAL VOC（pattern analysis, statistical modelling and computational learning visual object classes）2012数据集上比较了部分监督学习和弱监督学习的语义分割模型,并给出了监督学习模型和弱监督学习模型中的最优方法,以及对应的MIoU（mean intersection-over-union）。最后,指出了图像语义分割领域未来可能的热点方向。