CoralSeg: Learning coral segmentation from sparse annotations

Robotic advances and developments in sensors and acquisition systems facilitate the collection of survey data in remote and challenging scenarios. Semantic segmentation, which attempts to provide per‐pixel semantic labels, is an essential task when processing such data. Recent advances in deep learning approaches have boosted this task's performance. Unfortunately, these methods need large amounts of labeled data, which is usually a challenge in many domains. In many environmental monitoring instances, such as the coral reef example studied here, data labeling demands expert knowledge and is costly. Therefore, many data sets often present scarce and sparse image annotations or remain untouched in image libraries. This study proposes and validates an effective approach for learning semantic segmentation models from sparsely labeled data. Based on augmenting sparse annotations with the proposed adaptive superpixel segmentation propagation, we obtain similar results as if training with dense annotations, significantly reducing the labeling effort. We perform an in‐depth analysis of our labeling augmentation method as well as of different neural network architectures and loss functions for semantic segmentation. We demonstrate the effectiveness of our approach on publicly available data sets of different real domains, with the emphasis on underwater scenarios—specifically, coral reef semantic segmentation. We release new labeled data as well as an encoder trained on half a million coral reef images, which is shown to facilitate the generalization to new coral scenarios.     

Superparsing

This paper presents a simple and effective nonparametric approach to the problem of image parsing, or labeling image regions (in our case, superpixels produced by bottom-up segmentation) with their categories. This approach is based on lazy learning, and it can easily scale to datasets with tens of thousands of images and hundreds of labels. Given a test image, it first performs global scene-level matching against the training set, followed by superpixel-level matching and efficient Markov random field (MRF) optimization for incorporating neighborhood context. Our MRF setup can also compute a simultaneous labeling of image regions into semantic classes (e.g., tree, building, car) and geometric classes (sky, vertical, ground). Our system outperforms the state-of-the-art nonparametric method based on SIFT Flow on a dataset of 2,688 images and 33 labels. In addition, we report per-pixel rates on a larger dataset of 45,676 images and 232 labels. To our knowledge, this is the first complete evaluation of image parsing on a dataset of this size, and it establishes a new benchmark for the problem. Finally, we present an extension of our method to video sequences and report results on a video dataset with frames densely labeled at 1 Hz.     

基于深度自学习的图像哈希检索方法

基于监督学习的卷积神经网络被证明在图像识别的任务中具有强大的特征学习能力。然而,利用监督的深度学习方法进行图像检索,需要大量已标注的数据,否则很容易出现过拟合的问题。为了解决这个问题,提出了一种新颖的基于深度自学习的图像哈希检索方法。首先,通过无监督的自编码网络学习到一个具有判别性的特征表达函数,这种方法降低了学习的复杂性,让训练样本不需要依赖于有语义标注的图像,算法被迫在大量未标注的数据上学习更强健的特征。其次,为了加快检索速度,抛弃了传统利用欧氏距离计算相似性的方法,而使用感知哈希算法来进行相似性衡量。这两种技术的结合确保了在获得更好的特征表达的同时,获得了更快的检索速度。实验结果表明,提出的方法优于一些先进的图像检索方法。     

弱监督深度语义分割网络的多源遥感影像水体检测

目的 深度语义分割网络的优良性能高度依赖于大规模和高质量的像素级标签数据。在现实任务中,收集大规模、高质量的像素级水体标签数据将耗费大量人力物力。为了减少标注工作量,本文提出使用已有的公开水体覆盖产品来创建遥感影像对应的水体标签,然而已有的公开水体覆盖产品的空间分辨率低且存在一定错误。对此,提出采用弱监督深度学习方法训练深度语义分割网络。方法 在训练阶段,将原始数据集划分为多个互不重叠的子数据集,分别训练深度语义分割网络,并将训练得到的多个深度语义分割网络协同更新标签,然后利用更新后的标签重复前述过程,重新训练深度语义分割网络,多次迭代后可以获得好的深度语义分割网络。在测试阶段,多源遥感影像经多个代表不同视角的深度语义分割网络分别预测,然后投票产生最后的水体检测结果。结果 为了验证本文方法的有效性,基于原始多源遥感影像数据创建了一个面向水体检测的多源遥感影像数据集,并与基于传统的水体指数阈值分割法和基于低质量水体标签直接学习的深度语义分割网络进行比较,交并比（intersection-over-union,IoU）分别提升了5.5%和7.2%。结论 实验结果表明,本文方法具有收敛性,并且光学影像和合成孔径雷达（synthetic aperture radar,SAR）影像的融合有助于提高水体检测性能。在使用分辨率低、噪声多的水体标签进行训练的情况下,训练所得多视角模型的水体检测精度明显优于基于传统的水体指数阈值分割法和基于低质量水体标签直接学习的深度语义分割网络。     

基于边缘增强的遥感图像弱监督语义分割方法

随着弱监督学习被应用于遥感图像语义分割,大大降低了模型训练的数据成本。然而,由于监督信息不足,类激活图难以准确激活出遥感图像中不同尺度大小的目标,这使得基于类激活图获得的伪分割掩码边缘粗糙,从而导致最终的分割结果不准确。此外,大部分的弱监督语义分割方法都是基于可视化的两阶段方法,模型复杂繁琐。针对上述问题,设计了一种基于边缘增强的端到端弱监督语义分割网络。在特征空间边缘增强模块中,以自监督方式引导网络学习遥感图像中尺寸不一的目标,并且,细化伪分割掩码的边缘;在输出空间边缘增强模块中,通过端到端训练提升分割精度,同时降低模型训练的繁琐度。在ISPRS 2D数据集上的实验结果表明,该方法在仅使用图像级标签的情况下MIoU分别为57.72%和59.45%,与其他方法相比,效果较好。     

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。     

Training query filtering for semi-supervised learning to rank with pseudo labels

Semi-supervised learning is a machine learning paradigm that can be applied to create pseudo labels from unlabeled data for learning a ranking model, when there is only limited or no training examples available. However, the effectiveness of semi-supervised learning in information retrieval (IR) can be hindered by the low quality pseudo labels, hence the need for the training query filtering that removes the low quality queries. In this paper, we assume two application scenarios with respect to the availability of human labels. First, for applications without any labeled data available, a clustering-based approach is proposed to select the high quality training queries. This approach selects the training queries following the empirical observation that the relevant documents of high quality training queries are highly coherent. Second, for applications with limited labeled data available, a classification-based approach is proposed. This approach learns a weak classifier to predict the retrieval performance gain of a given training query by making use of query features. The queries with high performance gains are selected for the following transduction process to create the pseudo labels for learning to rank algorithms. Experimental results on the standard LETOR dataset show that our proposed approaches outperform the strong baselines.     

Fast and Scalable Mesh Superfacets

In the field of computer vision, the introduction of a low‐level preprocessing step to oversegment images into superpixels – relatively small regions whose boundaries agree with those of the semantic entities in the scene – has enabled advances in segmentation by reducing the number of elements to be labeled from hundreds of thousands, or millions, to a just few hundred. While some recent works in mesh processing have used an analogous oversegmentation, they were not intended to be general and have relied on graph cut techniques that do not scale to current mesh sizes. Here, we present an iterative superfacet algorithm and introduce adaptations of undersegmentation error and compactness, which are well‐motivated and principled metrics from the vision community. We demonstrate that our approach produces results comparable to those of the normalized cuts algorithm when evaluated on the Princeton Segmentation Benchmark, while requiring orders of magnitude less time and memory and easily scaling to, and enabling the processing of, much larger meshes.     

Face Annotation Using Transductive Kernel Fisher Discriminant

Face annotation in images and videos enjoys many potential applications in multimedia information retrieval. Face annotation usually requires many training data labeled by hand in order to build effective classifiers. This is particularly challenging when annotating faces on large-scale collections of media data, in which huge labeling efforts would be very expensive. As a result, traditional supervised face annotation methods often suffer from insufficient training data. To attack this challenge, in this paper, we propose a novel Transductive Kernel Fisher Discriminant (TKFD) scheme for face annotation, which outperforms traditional supervised annotation methods with few training data. The main idea of our approach is to solve the Fisher's discriminant using deformed kernels incorporating the information of both labeled and unlabeled data. To evaluate the effectiveness of our method, we have conducted extensive experiments on three types of multimedia testbeds: the FRGC benchmark face dataset, the Yahoo! web image collection, and the TRECVID video data collection. The experimental results show that our TKFD algorithm is more effective than traditional supervised approaches, especially when there are very few training data.     

Semi-supervised ranking for document retrieval

Ranking functions are an important component of information retrieval systems. Recently there has been a surge of research in the field of “learning to rank”, which aims at using labeled training data and machine learning algorithms to construct reliable ranking functions. Machine learning methods such as neural networks, support vector machines, and least squares have been successfully applied to ranking problems, and some are already being deployed in commercial search engines.Despite these successes, most algorithms to date construct ranking functions in a supervised learning setting, which assume that relevance labels are provided by human annotators prior to training the ranking function. Such methods may perform poorly when human relevance judgments are not available for a wide range of queries. In this paper, we examine whether additional unlabeled data, which is easy to obtain, can be used to improve supervised algorithms. In particular, we investigate the transductive setting, where the unlabeled data is equivalent to the test data.We propose a simple yet flexible transductive meta-algorithm: the key idea is to adapt the training procedure to each test list after observing the documents that need to be ranked. We investigate two instantiations of this general framework: The Feature Generation approach is based on discovering more salient features from the unlabeled test data and training a ranker on this test-dependent feature-set. The importance weighting approach is based on ideas in the domain adaptation literature, and works by re-weighting the training data to match the statistics of each test list. We demonstrate that both approaches improve over supervised algorithms on the TREC and OHSUMED tasks from the LETOR dataset.     

基于熵增强的无监督域适应遥感图像语义分割

为了实现利用有标注源数据获得在无标注目标数据上可用的遥感图像语义分割模型,提出了一种基于熵增强的域适应端到端语义分割方法.首先,为了充分利用遥感图像多尺度信息并且减少域之间传感器分辨率带来的域偏移,采用空洞空间金字塔池化模块作为分类器;其次,为了使无标注的目标域类别正确对应,使用了两个分类器进行协同训练;将像素点预测值的信息熵当做分类置信度的度量,将其作为对抗损失的权重,从而使训练能专注于难分类的像素,降低域偏移.在ISPRS(WGII/4)2D数据集上进行实验,所提方法相对于直接使用分割模型和使用传统对抗方法,mIoU分别提高了18％和12％.实验结果表明,所提方法在遥感图像域适应语义分割表现上优于直接使用分割模型或使用传统对抗域适应分割方法.     

近邻中心迭代策略的单标注视频行人重识别

为解决视频行人重识别数据集标注困难的问题,提出了基于单标注样本视频行人重识别的近邻中心迭代策略.该策略逐步利用伪标签视频片段迭代更新网络结构,以获得最佳的模型.针对预测无标签视频片段的伪标签准确率低的问题,提出了一种标签评估方法：每次训练后,将所选取的伪标签视频片段和有标签视频片段特征中每个类的中心点作为下一次训练中预测伪标签的度量中心点;同时提出基于交叉熵损失和在线实例匹配损失的损失控制策略,使得训练过程更加稳定,无标签数据的伪标签预测准确率更高.在MARS,DukeMTMC-VideoReID这两个大型数据集上的实验验证了该方法相比于最新的先进方法,在性能上得到非常好的提升.     

基于迁移学习的细粒度实体分类方法的研究

细粒度实体分类(Fine-grained entity type classification, FETC)旨在将文本中出现的实体映射到层次化的细分实体类别中. 近年来, 采用深度神经网络实现实体分类取得了很大进展. 但是, 训练一个具备精准识别度的神经网络模型需要足够数量的标注数据, 而细粒度实体分类的标注语料非常稀少, 如何在没有标注语料的领域进行实体分类成为难题. 针对缺少标注语料的实体分类任务, 本文提出了一种基于迁移学习的细粒度实体分类方法, 首先通过构建一个映射关系模型挖掘有标注语料的实体类别与无标注语料实体类别间的语义关系, 对无标注语料的每个实体类别, 构建其对应的有标注语料的类别映射集合. 然后, 构建双向长短期记忆(Bidirectional long short term memory, BiLSTM)模型, 将代表映射类别集的句子向量组合作为模型的输入用来训练无标注实体类别. 基于映射类别集中不同类别与对应的无标注类别的语义距离构建注意力机制, 从而实现实体分类器以识别未知实体分类. 实验证明, 我们的方法取得了较好的效果, 达到了在无任何标注语料前提下识别未知命名实体分类的目的.     

Using objective ground-truth labels created by multiple annotators for improved video classification: A comparative study

We address the problem of predicting category labels for unlabeled videos in a large video dataset by using a ground-truth set of objectively labeled videos that we have created. Large video databases like YouTube require that a user uploading a new video assign to it a category label from a prescribed set of labels. Such category labeling is likely to be corrupted by the subjective biases of the uploader. Despite their noisy nature, these subjective labels are frequently used as gold standard in algorithms for multimedia classification and retrieval. Our goal in this paper is NOT to propose yet another algorithm that predicts labels for unseen videos based on the subjective ground-truth. On the other hand, our goal is to demonstrate that the video classification performance can be improved if instead of using subjective labels, we first create an objectively labeled ground-truth set of videos and then train a classifier based on such a ground-truth so as to predict objective labels for the set of unlabeled videos.     

基于物体间支撑语义关系的室内场景彩色深度图像分割

场景理解是智能自主机器人领域的一个重要研究方向,而图像分割是场景理解的基础.但是,不完备的训练数据集,以及真实环境中的罕见情形,会导致在图像分割时存在先验知识不完备的情况,进而影响图像分割的效果.因此,提出在彩色深度(RGB–D)图像上使用抽象的支撑语义关系来解决多样的物体形态所面对的先验知识不完备问题.在先验知识不完备情况下,针对自底向上的图像分割过程中被过度分割出的物体块,首先对物体块间的支撑语义关系进行建模并计算其支撑概率,然后构造能够度量场景总体稳定性的能量函数,最后通过Swendsen-Wang割(SWC)随机图分割算法最小化该能量函数的值,将物体块间的支撑概率转化为强支撑语义关系并完成物体块合并,实现先验知识不完备情况下的图像分割.实验结果证明,结合支撑语义关系的图像分割能够在先验知识不完备的情况下,将同一物体被过度分割的部分重新合并起来,从而提升了图像分割的准确性.     

基于空间邻域信息的高光谱遥感影像半监督协同训练

针对tri_training协同训练算法在小样本的高光谱遥感影像半监督分类过程中,存在增选样本的误标记问题,提出一种基于空间邻域信息的半监督协同训练分类算法tri_training_SNI(tri_training based on Spatial Neighborhood Information)。首先利用分类器度量方法不一致度量和新提出的不一致精度度量从MLR(Multinomial Logistic Regression)、KNN(k-Nearest Neighbor)、ELM(Extreme Learning Machine)和RF(Random Forest)4个分类器中选择3分类性能差异性最大的3个分类器;然后在样本选择过程中,采用选择出来的3个分类器,在两个分类器分类结果相同的基础上,加入初始训练样本的8邻域信息进行未标记样本的二次筛选和标签的确定,提高了半监督学习的样本选择精度。通过对AVIRIS和ROSIS两景高光谱遥感影像进行分类实验,结果表明与传统的tri_training协同算法相比,该算法在分类精度方面有明显提高。     

Multi-label active learning by model guided distribution matching

Multi-label learning is an effective framework for learning with objects that have multiple semantic labels, and has been successfully applied into many real-world tasks. In contrast with traditional single-label learning, the cost of labeling a multi-label example is rather high, thus it becomes an important task to train an effectivemulti-label learning model with as few labeled examples as possible. Active learning, which actively selects the most valuable data to query their labels, is the most important approach to reduce labeling cost. In this paper, we propose a novel approach MADM for batch mode multi-label active learning. On one hand, MADM exploits representativeness and diversity in both the feature and label space by matching the distribution between labeled and unlabeled data. On the other hand, it tends to query predicted positive instances, which are expected to be more informative than negative ones. Experiments on benchmark datasets demonstrate that the proposed approach can reduce the labeling cost significantly.     

A new graph-based semi-supervised method for surface defect classification

Vision-based defect classification is an important technology to control the quality of product in manufacturing system. As it is very hard to obtain enough labeled samples for model training in the real-world production, the semi-supervised learning which learns from both labeled and unlabeled samples is more suitable for this task. However, the intra-class variations and the inter-class similarities of surface defect, named as the poor class separation, may cause the semi-supervised methods to perform poorly with small labeled samples. While graph-based methods, such as graph convolution network (GCN), can solve the problem well. Therefore, this paper proposes a new graph-based semi-supervised method, named as multiple micrographs graph convolutional network (MMGCN), for surface defect classification. Firstly, MMGCN performs graph convolution by constructing multiple micrographs instead of a large graph, and labels unlabeled samples by propagating label information from labeled samples to unlabeled samples in the micrographs to obtain multiple labels. Weighting the labels can obtain the final label, which can solve the limitations of computation complexity and practicality of original GCN. Secondly, MMGCN divides unlabeled dataset into multiple batches and sets an accuracy threshold. When the model accuracy reaches the threshold, the unlabeled datasets are labeled in batches. A famous case has been used to evaluate the performance of the proposed method. The experimental results demonstrate that the proposed MMGCN can achieve better computation complexity and practicality than GCN. And for accuracy, MMGCN can also obtain the best performance and the best class separation in the comparison with other semi-supervised surface defect classification methods.     

Ensemble learning from multiple information sources via label propagation and consensus

Many applications are facing the problem of learning from multiple information sources, where sources may be labeled or unlabeled, and information from multiple information sources may be beneficial but cannot be integrated into a single information source for learning. In this paper, we propose an ensemble learning method for different labeled and unlabeled sources. We first present two label propagation methods to infer the labels of training objects from unlabeled sources by making a full use of class label information from labeled sources and internal structure information from unlabeled sources, which are processes referred to as global consensus and local consensus, respectively. We then predict the labels of testing objects using the ensemble learning model of multiple information sources. Experimental results show that our method outperforms two baseline methods. Meanwhile, our method is more scalable for large information sources and is more robust for labeled sources with noisy data.     

Exploitation of unlabeled sequences in hidden Markov models

This paper presents a method for effectively using unlabeled sequential data in the learning of hidden Markov models (HMMs). With the conventional approach, class labels for unlabeled data are assigned deterministically by HMMs learned from labeled data. Such labeling often becomes unreliable when the number of labeled data is small. We propose an extended Baum-Welch (EBW) algorithm in which the labeling is undertaken probabilistically and iteratively so that the labeled and unlabeled data likelihoods are improved. Unlike the conventional approach, the EBW algorithm guarantees convergence to a local maximum of the likelihood. Experimental results on gesture data and speech data show that when labeled training data are scarce, by using unlabeled data, the EBW algorithm improves the classification performance of HMMs more robustly than the conventional naive labeling (NL) approach.