Abstract: Short video has been developing rapidly in recent years, and short video advertising has a promising prospect. However, the traditional advertisements are usually stiffly inserted into the videos, which are inefficient and always decrease users’ experience. This thesis proposes a systematic scheme for video object detection and content recommendation based on the deep learning model Faster R-CNN. This scheme will match the video contents to the displayed advertisements based on the principles of high correlation, precision and low interruption, thus obtains a balance between recommendation and user experience. Two system modes are available according to the video sources and network environments, named as Cloud Mode and Mobile Terminal Mode. The Cloud Mode is composed of a server, Content Delivery Network (CDN) and clients. The server will detect and recognize the contents of the CDN videos in advance, match them to corresponding advertisements by some recommendation algorithms and play the contents on the mobile Clients. The Mobile Terminal Mode mainly processes non-CDN resources such as some local videos, completes the tasks of object detection, recognition and content recommendation with limited computation ability. We apply the MobileNet model to improve the detection speed and accuracy, as well as to reduce memory footprint. To further increase efficiency and achieve real-time performance under the Mobile Terminal Mode, we implement joint compilation of Java and C++ code, adopt a self-developed player and cut down the object category based on the feedback system.

Key words: deep learning, object detection, content recommendation, Faster R-CNN, MobileNet