结合细粒度特征与深度卷积网络的手绘图检索

目的 传统的手绘图像检索方法主要集中在检索相同类别的图像，忽略了手绘图像的细粒度特征。对此，提出了一种新的结合细粒度特征与深度卷积网络的手绘图像检索方法，既注重通过深度跨域实现整体匹配，也实现细粒度细节匹配。方法 首先构建多通道混合卷积神经网络，对手绘图像和自然图像分别进行不同的处理；其次通过在网络中加入注意力模型来获取细粒度特征；最后将粗细特征融合，进行相似性度量，得到检索结果。结果 在不同的数据库上进行实验，与传统的尺度不变特征（SIFT）、方向梯度直方图（HOG）和深度手绘模型Deep SaN（sketch-a-net）、Deep 3DS（sketch）、Deep TSN（triplet sketch net）等5种基准方法进行比较，选取了Top-1和Top-10，在鞋子数据集上，本文方法Top-1正确率提升了12%，在椅子数据集上，本文方法Top-1正确率提升了11%，Top-10提升了3%，与传统的手绘检索方法相比，本文方法得到了更高的准确率。在实验中，本文方法通过手绘图像能在第1幅检索出绝大多数的目标图像，达到了实例级别手绘检索的目的。结论 提出了一种新的手绘图像检索方法，为手绘图像和自然图像的跨域检索提供了一种新思路，进行实例级别的手绘检索，与原有的方法相比，检索精度得到明显提升，证明了本文方法的可行性。     

A survey on freehand sketch recognition and retrieval

With the development of digital devices and pressure sensing equipment, research into freehand sketches from touch-screen interfaces has increased significantly in recent years. As such, we provide the first comprehensive survey of recognition tasks based on sketch generation, freehand sketch classification, sketch-based image retrieval (SBIR), fine-grained sketch-based image retrieval (FG-SBIR), and sketch-based 3D shape image retrieval. Specifically, SBIR and FG-SBIR were the main focus of the survey. Primary technologies and benchmark datasets related to all sketch-based recognition topics are also discussed, along with future trends for this promising technology.     

Combining pixel domain and compressed domain index for sketch based image retrieval

Sketch-based image retrieval (SBIR) lets one express a precise visual query with simple and widespread means. In the SBIR approaches, the challenge consists in representing the image dataset features in a structure that allows one to efficiently and effectively retrieve images in a scalable system. We put forward a sketch-based image retrieval solution where sketches and natural image contours are represented and compared, in both, the compressed-domain of wavelet and in the pixel domain. The query is efficiently performed in the wavelet domain, while effectiveness refinements are achieved using the pixel domain to verify the spatial consistency between the sketch strokes and the natural image contours. Also, we present an efficient scheme of inverted lists for sketch-based image retrieval using the compressed-domain of wavelets. Our proposal of indexing presents two main advantages, the amount of the data to compute the query is smaller than the traditional method while it presents a better effectiveness.     

基于跨模态自蒸馏的零样本草图检索

零样本草图检索将未见类的草图作为查询样本,用于检索未见类的图像。因此,这个任务同时面临两个挑战：草图和图像之间的模态差异以及可见类和未见类的不一致性。过去的方法通过将草图和图像投射到一个公共空间来消除模态差异,还通过利用语义嵌入（如词向量和词相似度）来弥合可见类和未见类的语义不一致。在本文中,我们提出了跨模态自蒸馏方法,从知识蒸馏的角度研究可泛化的特征,无需语义嵌入参与训练。具体而言,我们首先通过传统的知识蒸馏将预训练的图像识别网络的知识迁移到学生网络。然后,通过草图和图像的跨模态相关性,跨模态自蒸馏将上述知识间接地迁移到草图模态的识别上,提升草图特征的判别性和泛化性。为了进一步提升知识在草图模态内的集成和传播,我们进一步地提出草图自蒸馏。通过为数据学习辨别性的且泛化的特征,学生网络消除了模态差异和语义不一致性。我们在三个基准数据集,即Sketchy、TU-Berlin和QuickDraw,进行了广泛的实验,证明了我们提出的跨模态自蒸馏方法与当前方法相比较的优越性。     

包含跨域建模和深度融合网络的手绘草图检索

在手绘草图检索（sketch-based image retrieval,简称SBIR）领域,引入一种手绘草图的新型检索模型.手绘草图与自然图片之间存在巨大的差异性,这是因为,与自然图片相比,手绘草图展现出高度抽象的视觉表达,用现有的方法对手绘草图进行特征提取,其产生的特征描述子对于手绘草图的内容无法进行有效地拟合;对于相同的物体,不同的人群用手绘草图描述方式和表达也存在巨大的差距,这就使得手绘草图-自然图片的匹配更加困难;同时,将手绘草图与自然图片映射到相同视觉域的工作,也是一项具有困难的任务.所以,手绘草图检索技术是公认的比较有挑战性的任务.提出一种将手绘草图与自然图片在多个层次上映射到同一视觉域的策略来解决跨域的问题.同时,引入多层深度融合卷积神经网络（multi-layer deep fusion convolutional neural network）的框架来训练并获得手绘草图和自然彩色图片的多层特征表达.在Flickr15k图像数据库进行检索实验,实验结果显示,多层深度融合卷积网络学习到的特征的检索精度超过了现有的手工特征以及由自然图片或者手绘草图训练出来的卷积神经网络（convolutional neural network,简称CNN）的特征.     

基于自适应多类中心和半异构网络的三维模型草图检索

草图具有易于构建且不受语言、专业、年龄限制等优势,基于手绘草图的三维模型检索受到越来越多的关注.然而在三维模型草图检索任务中,三维模型具有复杂性,草图具有类内多样性,同时三维模型与草图之间又具有巨大的域间差异性,这些特点的相互作用严重影响检索的准确性.针对以上问题,提出了一种基于自适应多类中心和半异构网络的三维模型草图...     

Sketch-based 3D model retrieval by incorporating 2D-3D alignment

Sketch-based 3D model retrieval is very important for applications such as 3D modeling and recognition. In this paper, a sketch-based retrieval algorithm is proposed based on a 3D model feature named View Context and 2D relative shape context matching. To enhance the accuracy of 2D sketch-3D model correspondence as well as the retrieval performance, we propose to align a 3D model with a query 2D sketch before measuring their distance. First, we efficiently select some candidate views from a set of densely sampled views of the 3D model to align the sketch and the model based on their View Context similarities. Then, we compute the more accurate relative shape context distance between the sketch and every candidate view, and regard the minimum one as the sketch-model distance. To speed up retrieval, we precompute the View Context and relative shape context features of the sample views of all the 3D models in the database. Comparative and evaluative experiments based on hand-drawn and standard line drawing sketches demonstrate the effectiveness and robustness of our approach and it significantly outperforms several latest sketch-based retrieval algorithms.     

基于深度学习的草图检索方法研究进展

草图检索(SBIR)是基于内容的图像检索(CBIR)的扩展,是一种灵活便捷的目标图像检索方式,其研究的焦点是如何减少手绘草图域与自然图像域之间的域差。传统方法提取手工特征完成草图域与图像域之间的近似转换以减少域差,但该类方法无法有效拟合2个域内容,导致检索精度不高。深度学习方法依赖大量数据进行图像高维特征的提取,突破了传统方法的局限,已被证明可以有效解决跨域建模问题。研究聚焦于基于深度学习的草图检索方法,在深度特征提取模型、公开的数据测试集、粗粒度和细粒度检索、哈希技术和类别泛化等几个方面对草图检索的深度学习方法的相关研究工作进行了综述和评论。然后进行了实验比较研究,一方面,对现有3个公开的SBIR测试集Sketchy、TU-Berlin和QuickDraw进行适用性评估;另一方面,选取3个最新的SBIR深度学习模型GRLZS模型、SEM-PCYC模型和SAKE模型进行性能分析与比较。最后,对草图检索面临的挑战和未来研究方向进行了总结与展望。     

Discovering discriminative patches for free-hand sketch analysis

Since the ancient times, free-hand sketch has been widely used as an effective and convenient intermediate means to express human thoughts and highly diverse objects in reality. In recent years, a great quantity of researchers realized the significance of sketch and gradually focused on sketch-related problems, such as sketch-based image retrieval and recognition. Despite so many achievements, very few works concentrate on exploring the intrinsic factors which potentially influence the vivid degree of sketch. In this paper, we propose a weak supervised approach to discover the most discriminative patches for different categories of sketches, which perhaps grasp the key to a good free-hand sketch. In the beginning, we randomly extract tens of thousands of patches at multiple scales. After that, pyramid histogram of oriented gradient is calculated to represent these patches as an effective and uniform feature representation. To find the most discriminative patches for each class of sketches, we design an iterative detection process which combines cluster merging and discriminative ranking. The experimental results on the TU-Berlin sketch benchmark dataset demonstrate the effectiveness of the proposed method, as compared to other available approaches. Moreover, a reasonable analysis and discussion about good and bad sketches is provided based on the visual results.     

基于草图的人机交互技术研究进展

从草图识别和语义理解这两个方面对基于草图的人机交互技术的研究状况进行了分析和总结．对草图识别方法按其模式单元定义（笔划、图元、特征和组合图形）进行了分类和剖析;对草图语义理解所涉及的语义获取、语义解释和语义应用这三个关键问题及其解决方法进行了分析和阐述;并分别从基于草图的人机交互技术的几何模糊性、用户适应性和应用独旁性及其关系角度提出了这一领域的主要研究课题及其解决思路．     

基于草图纹理和形状特征融合的草图识别

人类具有很强的草图识别能力. 然而, 由于草图具有稀疏性和缺少细节的特点, 目前的深度学习模型在草图分类任务上仍然面临挑战. 目前的工作只是将草图看作灰度图像而忽略了不同草图类别间的形状表示差异. 提出一种端到端的手绘草图识别模型, 简称双模型融合网络, 它可以通过相互学习策略获取草图的纹理和形状信息. 具体地, 该模型由2个分支组成: 一个分支能够从图像表示(即原始草图)中自动提取纹理特征, 另一个分支能够从图形表示(即基于点的草图)中自动提取形状特征. 此外, 提出视觉注意一致性损失来度量2个分支之间视觉显著图的一致性, 这样可以保证2个分支关注相同的判别性区域. 最终将分类损失、类别一致性损失和视觉注意一致性损失结合完成双模型融合网络的优化. 在两个具有挑战性的数据集TU-Berlin数据集和Sketchy数据集上进行草图分类实验, 评估结果说明了双模型融合网络显著优于基准方法并达到最佳性能.     

基于可变形卷积的手绘图像检索

手绘图像仅包含简单线条轮廓, 与色彩、细节信息丰富的自然图像有着截然不同的特点. 然而目前的神经网络大多针对自然图像设计, 不能适应手绘图像稀疏性的特性. 针对此问题, 本文提出一种基于可变形卷积的手绘检索方法. 首先通过Berkerly边缘检测算法将自然图转化为边缘图, 消除域差异. 然后将卷积神经网络中的部分标准卷积替换为可变形卷积, 使网络能够充分关注手绘图轮廓信息. 最后分别将手绘图与边缘图输入网络并提取全连接层特征作为特征描述子进行检索. 在基准数据集Flickr15k上的实验结果表明, 本文方法与现有方法相比能够有效提高手绘图像检索精度.     

Query-by-Gesture: An Alternative Content-Based Image Retrieval Query Scheme

Current research on content-based image retrieval (CBIR) is centered on designing efficient query schemes in order to provide a user with effective mechanisms for image database search. Among representative CBIR query schemes, query-by-sketch has been one of the attractive query tools that are highly adaptive to user's subjectivity. However, query-by-sketch has a few limitations. That is, most sketch tools demand expertise in image processing or computer vision of the user to provide good enough sketches that can be used as query. Furthermore, sketching the exact shape of an object using a mouse can be a burden on the user. To overcome some of the limitations associated with query-by-sketch, we propose a new query method for CBIR, query-by-gesture, that does not require sketches, thereby minimizing user interaction. In our system, the user does not need to use a mouse to make a sketch. Instead, the user draws the shape of the object that heshe intends to search in front of a camera by hand. In addition, our query-by-gesture technique uses relevance feedback to interactively improve retrieval performance and allow progressive refinement of query results according to the user's specification. The efficacy of our proposed method is validated using images from the Corel-Photo CD.     

基于草图局部几何不变矩的图像检索方法

利用草图进行图像检索的难点在于对不同尺度、位置、旋转及形变图像的有效检索。为了更准确地识别并检索不同尺度、位置和旋转的图像,提出一种基于草图局部几何不变矩的图像检索方法（SBIRULGMI）。首先,利用图像的几何特征分别确定各图像的坐标系;然后,在生成的坐标系中对图像进行平均分块并计算各块的几何不变矩作为特征向量;接着,用改进的欧氏距离计算目标图像与数据库图像的相似度;最后,采用蚁群（ACO）算法对按照相似度排序后的检索结果进行优化。所提方法在MPEG-7 shape1 part B图像数据库的检索识别准确率比形状上下文（SC）、边缘分布直方图（EOH）、局部线性高波特征（GALIF）及MindFinder方法平均提高了17个百分点。实验结果表明该方法对不同平移、缩放和翻转的图像有较好的识别效果,对图像一定程度的旋转和形变具有更好的鲁棒性。     

Retrieving Aerial Scene Images with Learned Deep Image-Sketch Features

This paper investigates the problem of retrieving aerial scene images by using semantic sketches, since the state-of-the-art retrieval systems turn out to be invalid when there is no exemplar query aerial image available. However, due to the complex surface structures and huge variations of resolutions of aerial images, it is very challenging to retrieve aerial images with sketches and few studies have been devoted to this task. In this article, for the first time to our knowledge, we propose a framework to bridge the gap between sketches and aerial images. First, an aerial sketch-image database is collected, and the images and sketches it contains are augmented to various levels of details. We then train a multi-scale deep model by the new dataset. The fully-connected layers of the network in each scale are finally connected and used as cross-domain features, and the Euclidean distance is used to measure the cross-domain similarity between aerial images and sketches. Experiments on several commonly used aerial image datasets demonstrate the superiority of the proposed method compared with the traditional approaches.