面向互联网短视频数据的Logo检测技术研究

随着互联网技术发展，短视频数量呈爆炸式增长，给版权保护和品牌监管带来新的挑战，开展Logo自动检测是应对这一挑战的有效途径。受限于现有Logo数据集规模小、类别少以及Logo自身特点，现有目标检测方法用于Logo检测时效果不佳。本项目拟从Logo大规模数据集构建着手，对Logo的自动检测的关键技术展开研究：1）运用聚类技术对Logo数据进行分析，阐明基于先验的锚框选取机制；2）综合运用深度学习、注意力机制、优化理论等知识，构建针对大规模Logo数据的检测框架，提出多尺度特征提取和筛选方案，并通过对损失函数进行优化设计，解决Logo占比小在模型训练过程中造成的样本不均衡问题；3）基于一致性度量的预测回归设计，有效弥补现有目标检测方法中回归函数和评价指标之间存在的信息鸿沟。本课题有望突破Logo自动检测中的关键技术瓶颈，为互联网中短视频广告的有效检索、品牌监管和智能分析提供理论指导和技术支撑。

With the development of Internet technology, the number of short videos is growing explosively, which brings new challenges to copyright protection and brand supervision. Logo automatic detection is an effective way to deal with these challenges. Limited by the small scale, few categories of existing logo datasets and the characteristics of logo itself, the existing object detection methods are not effective in logo detection. This project intends to start from the construction of large-scale data set of logo, and study the key technologies of automatic detection of logo. Specifically, 1) The logo data is analyzed by employing clustering technology, and the anchor box selection mechanism is clarified based on prior information; 2) A detection framework for large-scale logo data is proposed by adopting the knowledge of deep learning, attention mechanism, optimization theory, etc. to achieve multi-scale feature extraction and selection. In the training, the problem of sample imbalance caused by the small proportion of logo is solved by optimizing the loss function; 3) The predictive regression design is proposed based on consistency measurement, which can effectively make up the information gap between regression function and evaluation index in the existing target detection methods. The implementation of the project is expected to break through the key technical bottleneck in logo detection as well as proivde the theoretical guidance and technical support for effective retrieval, brand supervision and intelligent analysis of short video advertising on the Internet.