紧致高效图像纹理特征表达与学习方法研究

Texture is an important characteristic of many types of images, ranging from multispectral remotely sensed images to microscopic images. Texture analysis plays a fundamental role image understanding, computer vision and pattern recognition. Real world applications pose a number of challenges to texture representation, such as the emerging big dimensionality, the prevalence of resources limited handheld devices, unconstrained imaging distortions, and the unavailability of large amount of annotated training data. These are prohibitively limiting factors for the use of most existing approaches such as large scale deep convolutional neural network based features. .The key challenges are addressed from three aspects. (i) This project intends to develop compact, efficient and discriminative texture specific Convolutional Neural Network (CNN) based representations (TexNets Family) by exploring innovative ideas from the fields of Sparse Representation (SR), Random Projection (RP), Compressed Sensing (CS) and Local Binary Pattern (LBP) methodology. (ii) There are many applications where only limited amounts of annotated training data can be available or collecting labeled training data is too expensive. In addition, it is not clear yet how general geometric invariances could be integrated into deep CNN networks. In order to handle such problems, this project intends to develop novel and robust LBP type texture descriptors (RoLBP Family) by exploring compact binary codes learning approaches that learns more discriminative and robust compact binary codes automatically from data. (iii) Novel methods for combining TexNets and RoLBP which aim to discover complementary and more powerful texture representations will be developed as well. .As a result of this project, theoretically, we can expect fruitful insights in DCNN understanding by providing formal connections among deep learning techniques, LBP methodology and the theories of SR, RP and CS. Algorithmically, we can expect the next generation texture methodology — revolutionary compact, efficient and discriminative texture representations that have low dimensionality, low computational complexity, robustness to imaging distortions and degradations, and adaptability to different types of problems and can be learned with moderate amounts of training data. Practically, novel results for face and facial expression recognition, and medical image analysis can also be expected.

纹理分析是计算机视觉和模式识别领域一个具有重要理论价值和广阔应用前景的研究课题。图像纹理表示面临一系列挑战，诸如高计算复杂度、资源受限的智能移动设备应用、特征高鲁棒性需求和高数据依赖性等，使得现有主流方法无法满足实际应用需求。.研究内容包括：研究紧致高效的基于深度卷积神经网络纹理特征表达方法（TexNets家族）；研究新颖的、鲁棒的、从数据中自动学习紧致二值纹理特征的方法（RoLBP家族）；研究融合TexNets家族和RoLBP家族获得具有互补性的更强大的纹理特征。.本课题预期取得理论创新与技术突破，促进纹理特征的广泛应用。理论上，建立深度学习、局部二值模式、以及稀疏表示和压缩感知理论之间联系，增强深度学习的可解释性；算法上，提出下一代紧致高效纹理特征表达与学习方法；应用上，将提出的方法应用于人脸识别、人脸（微）表情识别和医学图像处理。

纹理分析是计算机视觉和模式识别领域一个具有重要理论价值和广阔应用前景的研究课题。图像纹理表示面临一系列挑战，诸如高计算复杂度、资源受限的智能移动设备应用、特征高鲁棒性需求和高数据依赖性等，使得现有主流方法无法满足实际应用需求。研究内容包括：研究紧致高效的基于深度卷积神经网络纹理特征表达方法（TexNets家族）；研究新颖的、鲁棒的、从数据中自动学习紧致二值纹理特征的方法（RoLBP家族）；研究融合TexNets家族和RoLBP家族获得具有互补性的更强大的纹理特征。本课题预期取得理论创新与技术突破，促进纹理特征的广泛应用。理论上，建立深度学习、局部二值模式、以及稀疏表示和压缩感知理论之间联系，增强深度学习的可解释性；算法上，提出下一代紧致高效纹理特征表达与学习方法；应用上，将提出的方法应用于人脸识别、人脸（微）表情识别和医学图像处理。在本项目支持下，团队共发表/录用学术论文37篇，其中SCI检索论文28篇（IJCV/IEEE Transactions论文20篇，其中人工智能顶刊IEEE TPAMI 2篇、IJCV 3篇、IEEE TIP 3篇），自动化学报论文两篇，ICCV/ECCV/ACM MM等人工智能领域国际会议论文7篇。5项国家发明专利被授权，3项国家发明专利被受理。项目负责人刘丽研究员，入选国家“万人计划”青年拔尖人才（2022）、国家重点研发首席青年科学家（2021）、爱思唯尔中国高被引学者（2021）、国防科技大学第三批高层次创新人才科技领军人才培养对象（2022）、湖湘青年英才（2021）；2020年，项目负责人在视觉纹理信息紧致表示方面的研究，获中国电子学会自然科学一等奖。