基于像素级深度网络学习的互联网视频嵌入式标识识别

Superimposed logos represent rich semantic information of web video. Logo recognition, providing very important information for video content retrieval, classification, and content understanding, however, encounters many difficulties due to complex transforms and background influence in web video. To overtake these problems, firstly, we will design scheme to label logo based on temporal stability and spatial similarity, and propose a pixel-wise deep network learning framework, integrating the spatial and temporal cues, to locate the candidate logo region precisely, which can weaken the logo background influence effectively. For the candidate regions, based on local extrema detection，key points distribution characteristics and shape context of the key points, a scale invariant local descriptor will be designed to represent logo. Furthermore, the topological relationships will be integrated into the descriptor to improve discriminative power. Lastly, feature clustering and multi-level matching strategy via locality sensitive hashing will be proposed to achieve fast logo matching.

嵌入式视频标识包含丰富语义信息，检测与识别嵌入式标识对于视频内容检索、分类、内容理解均具有重要应用价值。针对互联网视频变化复杂、标识受背景干扰大等检测难题，首先研究基于时域稳定性和空域相似性的嵌入式标识的标注方法，进一步提出融合时空标注的像素级深度网络学习框架，准确获取嵌入式标识的前景区域，削弱背景影响；对候选标识区域，基于局部关键点检测、关键点的灰度变化特性分析、形状上下文，研究具有尺度不变性的局部描述子，进而融入拓扑关系，提高特征的区分能力。最后，利用局部敏感哈希索引建立特征聚类及分级匹配算法，实现标识的快速识别。

嵌入式视频标识包含丰富的语义信息，检测和识别嵌入式标识对于视频内容检索、分类、内容理解均具有重要的应用价值。针对互联网视频变化复杂、背景干扰等难题，研究了面向嵌入式标志检测的数据集获取、分类、标注、预处理，提出基于时域稳定性和空域相似性的标识标注方法。为提高模型的扩展能力，提出两阶段视频标志检测深度模型，其包含跨层像素级深度视频标志检测网络模型和标识识别模型，经过迭代优化、剪枝，构建标识检测原型系统PNET，并在实际工程中进行了验证，在30天的测试中准确率99.3%。研究表明，像素级深度网络的研究思路能够有效提升传统标识检测的精度。部分研究成果可以拓展到其他检测任务领域，包括显著性计算模型优化、SSD模型优化及应用研究、基于GAN的数据扩增技术。