Image Annotation by Propagating Labels from Semantic Neighbourhoods

Automatic image annotation aims at predicting a set of semantic labels for an image. Because of large annotation vocabulary, there exist large variations in the number of images corresponding to different labels (“class-imbalance”). Additionally, due to the limitations of human annotation, several images are not annotated with all the relevant labels (“incomplete-labelling”). These two issues affect the performance of most of the existing image annotation models. In this work, we propose 2-pass k-nearest neighbour (2PKNN) algorithm. It is a two-step variant of the classical k-nearest neighbour algorithm, that tries to address these issues in the image annotation task. The first step of 2PKNN uses “image-to-label” similarities, while the second step uses “image-to-image” similarities, thus combining the benefits of both. We also propose a metric learning framework over 2PKNN. This is done in a large margin set-up by generalizing a well-known (single-label) classification metric learning algorithm for multi-label data. In addition to the features provided by Guillaumin et al. (2009) that are used by almost all the recent image annotation methods, we benchmark using new features that include features extracted from a generic convolutional neural network model and those computed using modern encoding techniques. We also learn linear and kernelized cross-modal embeddings over different feature combinations to reduce semantic gap between visual features and textual labels. Extensive evaluations on four image annotation datasets (Corel-5K, ESP-Game, IAPR-TC12 and MIRFlickr-25K) demonstrate that our method achieves promising results, and establishes a new state-of-the-art on the prevailing image annotation datasets.     

Compressed labeling on distilled labelsets for multi-label learning

Directly applying single-label classification methods to the multi-label learning problems substantially limits both the performance and speed due to the imbalance, dependence and high dimensionality of the given label matrix. Existing methods either ignore these three problems or reduce one with the price of aggravating another. In this paper, we propose a {0,1} label matrix compression and recovery method termed ??compressed labeling (CL)?? to simultaneously solve or at least reduce these three problems. CL first compresses the original label matrix to improve balance and independence by preserving the signs of its Gaussian random projections. Afterward, we directly utilize popular binary classification methods (e.g., support vector machines) for each new label. A fast recovery algorithm is developed to recover the original labels from the predicted new labels. In the recovery algorithm, a ??labelset distilling method?? is designed to extract distilled labelsets (DLs), i.e., the frequently appeared label subsets from the original labels via recursive clustering and subtraction. Given a distilled and an original label vector, we discover that the signs of their random projections have an explicit joint distribution that can be quickly computed from a geometric inference. Based on this observation, the original label vector is exactly determined after performing a series of Kullback-Leibler divergence based hypothesis tests on the distribution about the new labels. CL significantly improves the balance of?the training samples and reduces the dependence between different labels. Moreover, it accelerates the learning process by training fewer binary classifiers for compressed labels, and makes use of label dependence via DLs based tests. Theoretically, we prove the recovery bounds of CL which verifies the effectiveness of CL for label compression and multi-label classification performance improvement brought by label correlations preserved in DLs. We show the effectiveness, efficiency and robustness of CL via 5 groups of experiments on 21 datasets from text classification, image annotation, scene classification, music categorization, genomics and web page classification.     

基于分类融合和关联规则挖掘的图像语义标注

图像语义自动标注问题是现阶段一个具有挑战性的难题。在跨媒体相关模型基础上,提出了融合图像类别信息的图像语义标注新方法,并利用关联规则挖掘算法改善标注结果。首先对图像进行低层特征提取,用“视觉词袋”描述图像;然后对图像特征分别进行K-means聚类和基于支持向量机的多类别分类,得到图像相似性关系和类别信息;计算语义标签和图像之间的概率关系,并将图像类别信息作为权重融合到标签的统计概率中,得到候选标注词集;最后以候选标注词概率为依据,利用改善的关联规则挖掘算法挖掘文本关联度,并对候选标注词集进行等频离散化处理,从而得到最终标注结果。在图像集Corel上进行的标注实验取得了较为理想的标注结果。     

融合深度特征和语义邻域的自动图像标注*

传统图像标注方法中人工选取特征费时费力,传统标签传播算法忽视语义近邻,导致视觉相似而语义不相似,影响标注效果.针对上述问题,文中提出融合深度特征和语义邻域的自动图像标注方法.首先构建基于深度卷积神经网络的统一、自适应深度特征提取框架,然后对训练集划分语义组并建立待标注图像的邻域图像集,最后根据视觉距离计算邻域图像各标签的贡献值并排序得到标注关键词.在基准数据集上实验表明,相比传统人工综合特征,文中提出的深度特征维数更低,效果更好.文中方法改善传统视觉近邻标注方法中的视觉相似而语义不相似的问题,有效提升准确率和准确预测的标签总数.     

Learning multi-task local metrics for image annotation

The goal of image annotation is to automatically assign a set of textual labels to an image to describe the visual contents thereof. Recently, with the rapid increase in the number of web images, nearest neighbor (NN) based methods have become more attractive and have shown exciting results for image annotation. One of the key challenges of these methods is to define an appropriate similarity measure between images for neighbor selection. Several distance metric learning (DML) algorithms derived from traditional image classification problems have been applied to annotation tasks. However, a fundamental limitation of applying DML to image annotation is that it learns a single global distance metric over the entire image collection and measures the distance between image pairs in the image-level. For multi-label annotation problems, it may be more reasonable to measure similarity of image pairs in the label-level. In this paper, we develop a novel label prediction scheme utilizing multiple label-specific local metrics for label-level similarity measure, and propose two different local metric learning methods in a multi-task learning (MTL) framework. Extensive experimental results on two challenging annotation datasets demonstrate that 1) utilizing multiple local distance metrics to learn label-level distances is superior to using a single global metric in label prediction, and 2) the proposed methods using the MTL framework to learn multiple local metrics simultaneously can model the commonalities of labels, thereby facilitating label prediction results to achieve state-of-the-art annotation performance.     

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.     

基于迁移学习与多标签平滑策略的图像自动标注

针对图像标注数据集标签分布不平衡问题,提出了基于标签平滑策略的多标签平滑单元（MLSU）。MLSU在网络模型训练过程中自动平滑数据集中的高频标签,使网络适当提升了低频标签的输出值,从而提升了低频标注词的标注性能。为解决图像标注数据集样本数量不足造成网络过拟合的问题,提出了基于迁移学习的卷积神经网络（CNN）模型。首先利用互联网上的大型公共图像数据集对深度网络进行预训练,然后利用目标数据集对网络参数进行微调,构建了一个多标签平滑卷积神经网络模型（CNN-MLSU）。分别在Corel5K和IAPR TC-12图像标注数据集上进行实验,在Corel5K数据集上,CNN-MLSU较卷积神经网络回归方法（CNN-R）的平均准确率与平均召回率分别提升了5个百分点和8个百分点;在IAPR TC-12数据集上,CNN-MLSU较两场K最邻近模型（2PKNN_ML）的平均召回率提升了6个百分点。实验结果表明,基于迁移学习的CNN-MLSU方法能有效地预防网络过拟合,同时提升了低频词的标注效果。     

基于EM的非监督图像多标签区域标定算法

提出一个基于EM迭代的非监督图像多标签区域标定算法,它能够非常有效地将基于全图的标签自动标定到图像的对应局部区域上。首先对所有图像进行SIFT特征点的密集采样,然后对所有的SIFT特征点进行K-m eans聚类,获得词典,再构造EM迭代过程计算出每幅图像中每个标签对每个存在WORD的置信度,最后选择那些置信度较高的WORD,确定每幅图像中每个标签置信度最高的对应区域。实验表明,在样本数据充分的情况下,该算法在解决非监督自动标定、标签表观的多样性以及多标签等问题上都取得了不错的效果。     

基于低秩稀疏分解优化的图像标签完备

大量上传的网络图像因用户语义标注的随意性,造成了图像标签的不完备,大大降低了图像检索的效率.低秩稀疏是一种有效降低数据噪声的方法.为提高图像语义标签完备的准确度,提出一种基于低秩稀疏分解优化(LRSDO)的图像标签完备方法.首先结合待完备图像的视觉特征和语义搜索其近邻图像集;然后通过低秩稀疏分解模型获得其视觉特征与语义之间的映射关系,并以此预测该图像的候选标签;最后使用面向个体的标签共现频率方法对候选标签进行去噪优化,进而实现对其更加准确的自动图像标签完备.在基准数据集Corel5K和真实数据集Flickr30Concepts上进行了实验,结果表明,该方法在图像标签完备的平均准确率,平均召回率和覆盖率上均表现出更优的性能.     

基于标记特定特征和相关性的ML-KNN改进算法

目前大部分已经存在的多标记学习算法在模型训练过程中所采用的共同策略是基于相同的标记属性特征集合预测所有标记类别.但这种思路并未对每个标记所独有的标记特征进行考虑.在标记空间中,这种标记特定的属性特征对于区分其它类别标记和描述自身特性是非常有帮助的信息.针对这一问题,本文提出了基于标记特定特征和相关性的ML-KNN改进算法MLF-KNN.不同于之前的多标记算法直接在原始训练数据集上进行操作,而是首先对训练数据集进行预处理,为每一种标记类别构造其特征属性,在得到的标记属性空间上进一步构造L₁-范数并进行优化从而引入标记之间的相关性,最后使用改进后的ML-KNN算法进行预测分类.实验结果表明,在公开数据集image和yeast上,本文提出的算法MLF-KNN分类性能优于ML-KNN,同时与其它另外3种多标记学习算法相比也表现出一定的优越性.     

A two-level model for automatic image annotation

Image automatic annotation is a significant and challenging problem in pattern recognition and computer vision. Current image annotation models almost used all the training images to estimate joint generation probabilities between images and keywords, which would inevitably bring a lot of irrelevant images. To solve the above problem, we propose a hierarchical image annotation model which combines advantages of discriminative model and generative model. In first annotation layer, discriminative model is used to assign topic annotations to unlabeled images, and then relevant image set corresponding to each unlabeled image is obtained. In second annotation layer, we propose a keywords-oriented method to establish links between images and keywords, and then our iterative algorithm is used to expand relevant image sets. Candidate labels will be given higher weights by using our method based on visual keywords. Finally, generative model is used to assign detailed annotations to unlabeled images on expanded relevant image sets. Experiments conducted on Corel 5K datasets verify the effectiveness of our hierarchical image annotation model.     

Evaluation of Localized Semantics: Data, Methodology, and Experiments

We present a new data set of 1014 images with manual segmentations and semantic labels for each segment, together with a methodology for using this kind of data for recognition evaluation. The images and segmentations are from the UCB segmentation benchmark database (Martin et al., in International conference on computer vision, vol. II, pp. 416–421, 2001). The database is extended by manually labeling each segment with its most specific semantic concept in WordNet (Miller et al., in Int. J. Lexicogr. 3(4):235–244, 1990). The evaluation methodology establishes protocols for mapping algorithm specific localization (e.g., segmentations) to our data, handling synonyms, scoring matches at different levels of specificity, dealing with vocabularies with sense ambiguity (the usual case), and handling ground truth regions with multiple labels. Given these protocols, we develop two evaluation approaches. The first measures the range of semantics that an algorithm can recognize, and the second measures the frequency that an algorithm recognizes semantics correctly. The data, the image labeling tool, and programs implementing our evaluation strategy are all available on-line (kobus.ca//research/data/IJCV_2007). We apply this infrastructure to evaluate four algorithms which learn to label image regions from weakly labeled data. The algorithms tested include two variants of multiple instance learning (MIL), and two generative multi-modal mixture models. These experiments are on a significantly larger scale than previously reported, especially in the case of MIL methods. More specifically, we used training data sets up to 37,000 images and training vocabularies of up to 650 words. We found that one of the mixture models performed best on image annotation and the frequency correct measure, and that variants of MIL gave the best semantic range performance. We were able to substantively improve the performance of MIL methods on the other tasks (image annotation and frequency correct region labeling) by providing an appropriate prior.     

Implicit relative attribute enabled cross-modality hashing for face image-video retrieval

Face image-video retrieval refers to retrieving videos of a specific person with image query or searching face images of one person by using a video clip query. It has attracted much attention for broad applications like suspect tracking and identifying. This paper proposes a novel implicit relative attribute enabled cross-modality hashing (IRAH) method for large-scale face image-video retrieval. To cope with large-scale data, the proposed IRAH method facilitates fast cross-modality retrieval through embedding two entirely heterogeneous spaces, i.e., face images in Euclidean space and face videos on a Riemannian manifold, into a unified compact Hamming space. In order to resolve the semantic gap, IRAH maps the original low-level kernelized features to discriminative high-level implicit relative attributes. Therefore, the retrieval accuracy can be improved by leveraging both the label information across different modalities and the semantic structure obtained from the implicit relative attributes in each modality. To evaluate the proposed method, we conduct extensive experiments on two publicly available databases, i.e., the Big Bang Theory (BBT) and Buffy the Vampire Slayer (BVS). The experimental results demonstrate the superiority of the proposed method over different state-of-the-art cross-modality hashing methods. The performance gains are especially significant in the case that the hash code length is 8 bits, up to 12% improvements over the second best method among tested methods.

     

Neural ranking for automatic image annotation

Automatic image annotation aims to predict labels for images according to their semantic contents and has become a research focus in computer vision, as it helps people to edit, retrieve and understand large image collections. In the last decades, researchers have proposed many approaches to solve this task and achieved remarkable performance on several standard image datasets. In this paper, we propose a novel learning to rank approach to address image auto-annotation problem. Unlike typical learning to rank algorithms for image auto-annotation which directly rank annotations for image, our approach consists of two phases. In the first phase, neural ranking models are trained to rank image’s semantic neighbors. Then nearest-neighbor based models propagate annotations from these semantic neighbors to the image. Thus our approach integrates learning to rank algorithms and nearest-neighbor based models, including TagProp and 2PKNN, and inherits their advantages. Experimental results show that our method achieves better or comparable performance compared with the state-of-the-art methods on four challenging benchmarks including Corel5K, ESP Games, IAPR TC-12 and NUS-WIDE.

     

基于连续预测的半监督学习图像语义标注

为了在图像底层特征与高层语义之间建立关系,提高图像自动标注的精确度,结合基于图学习的方法和基于分类的标注算法,提出了基于连续预测的半监督学习图像语义标注的方法,并对该方法的复杂度进行分析。该方法利用标签数据提供的信息和标签事例与无标签事例之间的关系,根据邻接点(事例)属于同一个类的事实,构建K邻近图。用一个基于图的分类器,通过核函数有效地计算邻接信息。在建立图的基础上,把经过划分后的样本节点集通过基于连续预测的多标签半监督学习方法进行标签传递。实验表明,提出的算法在图像标注中的标注词的平均查准率、平均查全率方面有显著的提高。     

Leaf Bagging: A novel meta heuristic optimization based framework for leaf identification

Automated plant recognition based on leaf images is a challenging task among the researchers from several fields. This task requires distinguishing features derived from leaf images for assigning class label to a leaf image. There are several methods in literature for extracting such distinguishing features. In this paper, we propose a novel automated framework for leaf identification. The proposed framework works in multiple phases i.e. pre-processing, feature extraction, classification using bagging approach. Initially, leaf images are pre-processed using image processing operations such as boundary extraction and cropping. In the feature extraction phase, popular nature inspired optimization algorithms viz. Spider Monkey Optimization (SMO), Particle Swarm Optimization (PSO) and Gray Wolf Optimization (GWO) have been exploited for reducing the dimensionality of features. In the last phase, a leaf image is classified by multiple classifiers and then output of these classifiers is combined using majority voting. The effectiveness of the proposed framework is established based on the experimental results obtained on three datasets i.e. Flavia, Swedish and self-collected leaf images. On all the datasets, it has been observed that the classification accuracy of the proposed method is better than the individual classifiers. Furthermore, the classification accuracy for the proposed approach is comparable to deep learning based method on the Flavia dataset.

     

Label distribution learning for scene text detection

Recently, segmentation-based scene text detection has drawn a wide research interest due to its flexibility in describing scene text instance of arbitrary shapes such as curved texts. However, existing methods usually need complex post-processing stages to process ambiguous labels, i.e., the labels of the pixels near the text boundary, which may belong to the text or background. In this paper, we present a framework for segmentation-based scene text detection by learning from ambiguous labels. We use the label distribution learning method to process the label ambiguity of text annotation, which achieves a good performance without using additional post-processing stage. Experiments on benchmark datasets demonstrate that our method produces better results than state-of-the-art methods for segmentation-based scene text detection.     

基于标签集相关性学习的大规模网络图像在线标注

传统的网络图像标注方法忽视了标签集整体相关性对标注结果的影响，导致标签集整体相关性缺乏和语义冗余. 为了解决上述问题，提出了一种基于标签集相关性学习的大规模网络图像在线语义标注方法. 给出了标签集对图像相关性和标签集内部相关性的概率估计算法，将上述约束形成一个优化问题，采用贪心搜索策略获取近似最优解，找到能合理地平衡上述因素的标签集，并针对大规模图像集和概念集进行了优化. 真实环境下大规模网络图像集上的测试表明，相比于目前的代表性网络图像标注方法，该方法获得的标签集能够更好的描述图像语义，性能提升明显.     

Label noise correction and application in crowdsourcing

The important task of correcting label noise is addressed infrequently in literature. The difficulty of developing a robust label correction algorithm leads to this silence concerning label correction. To break the silence, we propose two algorithms to correct label noise. One utilizes self-training to re-label noise, called Self-Training Correction (STC). Another is a clustering-based method, which groups instances together to infer their ground-truth labels, called Cluster-based Correction (CC). We also adapt an algorithm from previous work, a consensus-based method called Polishing that consults with an ensemble of classifiers to change the values of attributes and labels. We simplify Polishing such that it only alters labels of instances, and call it Polishing Labels (PL). We experimentally compare our novel methods with Polishing Labels by examining their improvements on the label qualities, model qualities, and AUC metrics of binary and multi-class data sets under different noise levels. Our experimental results demonstrate that CC significantly improves label qualities, model qualities, and AUC metrics consistently. We further investigate how these three noise correction algorithms improve the data quality, in terms of label accuracy, in the context of image labeling in crowdsourcing. First, we look at three consensus methods for inferring a ground-truth label from the multiple noisy labels obtained from crowdsourcing, i.e., Majority Voting (MV), Dawid Skene (DS), and KOS. We then apply the three noise correction methods to correct labels inferred by these consensus methods. Our experimental results show that the noise correction methods improve the labeling quality significantly. As an overall result of our experiments, we conclude that CC performs the best. Our research has illustrated the viability of implementing noise correction as another line of defense against labeling error, especially in a crowdsourcing setting. Furthermore, it presents the feasibility of the automation of an otherwise manual process of analyzing a data set, and correcting and cleaning the instances, an expensive and time-consuming task.