基于图像属性和深度学习的大规模物体检测研究与应用

Along with the development of modern science and technology, the number of images and videos increase rapidly. How to accurately and efficiently detect human faces, pedestrians, and objects becomes an important topic in the era of big data. Object detection is a fundamental technology which helps extract beneficial knowledge from internet, builds large scale surveillance system, and improves the next generation artificial intelligence such as robotics and drone. However, the conventional object detection methods face with new challenges in the big data era. For example, first, unconstrained and large scale image and video data often have large and complex variants, including crowd, occlusion, low resolution, and viewpoints. Second, the traditional object detection methods often used supervised learning, which needs a large number of annotated data. Nevertheless, data labeling in large dataset costs a lot of resources. Third, efficiency is an important issue in practical applications. To solve the above problems, this project combines attributes and deep learning to jointly model and train different key components of object detection. We are dealing with the following key issues: 1) large scale object detection in complex scenes; 2) weakly- and semi-supervised deep learning; and 3) improvement of the efficiency of deep models. This project adopts deep learning and improves large scale object detection, making it possible to be applied to object detection in real-world applications. In the aspect of algorithm, we will design weakly and semi-supervised deep learning methods and deep model compression. In the aspect of application, we will apply our deep models to large scale face detection, pedestrian detection, and object detection.

随着科技的高速发展，图像与视频数据持续增加并成为主要信息载体。如何准确高效的检测人脸、行人与常见物体已成为大数据时代一个迫切需要解决的问题。物体检测是抽取互联网有用信息、排除劣质信息的技术保障，是构建“平安城市”视频监控网络的基础，是下一代人工智能技术视觉系统如机器人和无人驾驶等的重要组成部分。大规模、非受控复杂图像与视频数据给传统物体检测技术带来巨大挑战。本项目使用图像属性和深度学习对物体检测流程进行统一建模与优化。拟解决的关键问题包括：1）复杂场景的大规模物体检测方法；2）弱监督与半监督深度学习方法；3）深度学习模型的时间效率优化。预期成果：算法方面，完成弱监督与半监督深度学习建模、优化和分析方法。并探索深度网络模型的压缩方法；应用方面，结合算法研究，搭建复杂场景下人脸、行人与常见物体检测流程。本项目的研究对推动物体检测技术在复杂环境下应用有着重要意义。

本项目开展在大规模图像与视频中，如何准确高效的检测人脸、行人与常见物体。这些问题已成为大数据时代一个迫切需要解决的关键问题。它对科技进步与社会发展有着至关重要的作用。本项目以深度学习为基础，为深度学习物体检测、属性识别、模型压缩、模型结构搜索开发新的框架，并以此为核心，研发在物体识别与检测、分割等领域的具体应用。经过全体成员共同努力，取得成果：.1. 深度学习理论研究：利用先进的机器学习理论，为深度学习训练优化和结构搜索过程建立联系。成果发表与ICML 17， IJCAI17，和BigData18等顶级国际会议。.2. 物体检测、属性识别应用研究：构建大规模高影响力的数据集3项，包括CelebA，DeepFashion, WIDER Face。其中CelebA谷歌学术引用超过1000次。三个数据库总共引用超过1500次。.3. 完成CCF-A类论文19篇。其中包括人工智能顶级会议AAAI和IJCAI论文2篇，机器学习顶级会议ICML和NIPS论文2篇，与计算机视觉顶级会议CVPR和ICCV论文9篇。上述成果参加国际会议，并被邀请做口头报告6次。另外，本项目相关工作被邀请投稿至大数据顶级会议IEEE Big Data正式长文论文2篇，并做口头报告。本项目取得具有国际影响力的学术成果，在这一领域继续保持国际领先水平。.4. 申请发明专利3项。.5. 培养研究生4名。..通过本项目科研，.1. 项目期间，项目负责人获得深圳市海外高层次人才（C 类）、深圳市南山区领航人才（C 类）。.2. 在2016年人工智能顶级会议AAAI Conference on Artificial Intelligence，Face Model Compression by Distilling Knowledge from Neurons发表的论文接收为口头报告。接收率小于5%（投稿数大于4000）。.3. 在2017年机器学习顶级会议The Thirty-fourth International Conference on Machine Learning (ICML)发表唯一作者论文Learning Deep Architectures via Generalized Whitened Neural Networks。被邀请做口头报告。