噪声环境下基于深度学习的图像自动标注方法研究

Automatic image annotation is the foreland problem in the fields of pattern recognition and computer vision. Currently, there are kinds of noises in the most real image annotation environment, while the systematic studies of automatic image annotation in noise environment had not yet appeared. Accordingly, aiming at the automatic image annotation in noise environment, the following three aspects research would be carried out: (1) In view of the traditional shallow learning methods lacking enough modeling and generalization capabilities in handling complex problems, a unified automatic image annotation framework based on Prototype-based Sparse Denoising AutoEncoders is proposed. The framework is based on deep learning. Sparse rules constraints were added on the basis of Denoising AutoEncoders, and proposed unified image concept prototypes by integration of single-instance representation and multi-instances representation were used to restrict the feature learning processes of Sparse Denoising AutoEncoders. The goals of reducing noise layer by layer and improving generalization capabilities of image annotation would be reached by building the deep network. The framework does not depend on specific automatic image annotation methods, and the framework has the capacities of learning objects’ intrinsic features under noisy environments. (2) Aiming at the emerging large-scale images in the new field, an image annotation method based on Co-clustering (IAC) is proposed. IAC is based on transfer learning, which can effectively transfer image annotation to large-scale new image database by using small number of annotation images from corresponding new fields and large number of auxiliary training data together. (3) Proposed group-based visual context method is used to refine the image annotation results. In this project, effective solutions of automatic image annotation in noisy environment were proposed to enhance the performance of automatic image annotation in real environment. So this project has important theoretical significances and practical values.

图像自动标注是模式识别与计算机视觉等领域的前沿课题。目前，真实图像标注环境大多含有各类噪声，而现有几乎还未出现围绕噪声环境的图像标注研究，因此，本课题针对噪声环境下的图像自动标注，开展以下三方面研究：1)考虑到传统浅层学习方法对复杂问题的建模与泛化能力不足，提出基于原型的稀疏降噪自动编码机统一图像自动标注框架，该框架基于深度学习，通过构建深度网络达到逐层降低噪声与提升标注泛化能力，同时该框架不依赖于具体某类标注方法，能够在噪声环境下学习目标的本质特征；2)针对不断出现的大规模新领域图像，提出具有迁移能力的互聚类图像自动标注算法，该方法通过少量新领域标注图像与大量辅助训练数据将标注信息有效迁移到大规模新领域图像库中；3)利用基于群组的视觉上下文图像标注改善模型对标注结果进行改善。本课题针对噪声环境下的图像自动标注提出有效解决方案，提升真实环境下图像标注的性能，具有重要理论意义和实际应用价值。

图像自动标注是模式识别与计算机视觉等领域的前沿课题。目前，真实图像标注环境大多含有各类噪声，本课题针对噪声环境下的图像自动标注，开展的主要研究工作包括：提出了基于线性栈式自动编码器（L-SAE）的自动图像标注模型，并在此模型基础上提出一种分组强化训练L-BSAE子模型的线性鲁棒平衡栈式自动编码器算法（L-RBSAE），解决了SAE模型训练中小型数据速度上的不足；提出了基于非线性栈式自动编码器（NL-SAE）的自动图像标注模型，从一定程度上解决了传统浅层机器学习算法在处理复杂分类问题时存在泛化能力不足的问题；提出一种增强训练中低频标签的非线性平衡栈式自动编码器（NL-BSAE），并在此模型基础上提出一种分组强化训练NL-BSAE子模型的非线性鲁棒平衡栈式自动编码器算法（NL-RBSAE），提升了图像自动标注模型处理不平衡数据的能力；提出了融合深度特征和语义邻域的自动图像标注方法，解决了传统标签传播算法忽视语义近邻而影响标注效果等问题。提出了多尺度特征融合提取方法，该方法结合图像的全局特征信息和局部特征信息，对图像的曝光，障碍物遮挡，以及放大和缩小等问题均具有较强的鲁棒性。提出了鲁棒性增量极限学习机图像自动标注方法，该方法结合了特征抑制和增量反馈的思想，能够在一定程度上提高标注的速度。提出了基于低秩表示的上下文正则化的运动目标检测模型，解决了传统方法需要针对每一场景逐一进行参数训练等问题；提出基于稀疏模型的多场景串流视频检测模型，将运动目标检测问题转化为低秩表示下的连续性离群点检测问题，提高了多场景串流视频的检测准确率，有效解决了传统检测算法无法适用于多场景串流视频检测的问题；提出了鲁棒均衡策略的区域全卷积网络（BRR-FCN），通过不同框架和模型相融合的方式进一步提升模型的效果。本项目的理论研究还应用于人脸检测与识别、行人检测与识别、人体姿态估计与属性识别、交通要素检测与目标跟踪等实际场景。