以支援向量機硬體為基礎的影片分段界線偵測系統設計與實作

Abstract

Structural analysis of video is an essential step to automatic video content analysis. In a general video structure, shot is the most basic unit, and it must be determined before any multimedia content-based retrieval applications. However, despite the rich research efforts in software shot detection system, there is a lack of shot detection system that is designed in hardware. In this thesis, we proposed a shot boundary detection system using hardwarebased Support-Vector-Machine(SVM). Our system can detect both cut shots and gradual shots. We optimized the feature extraction process based on global color histogram with a pipelined architecture to save memory and increase the overall speed. The class imbalance problem in shot detection is solved by using random pseudo-sampling at the SVM training stage. The digital hardware SVM is designed using a fully-parallel pipelined architecture and is highly configurable on vector dimensions. Our data wordlength only used 4 bits signed integer plus one bit decimal in fixed number, while the detection accuracy is competitive comparing with floating point software. Our SVM classifier presented here can run a speed up to 251.62MHZ on Xilinx Virtex IV XC4VSX35 FPGA, and the video processing on our shot detection system can achieve 128 fps on PC. That makes our system met the real-time constraints, and detect shots in a continous video stream such as live-TV news or sports game.