Shot Boundary Detection Using Adaptive Threshold And Fitting Feature

Shot boundary detection technology plays a very important role in the area of digital video process. Researchers and technicians have been focused on shot boundary detection ever since its first emergence. This technology is intended to detect transitions in video, including hard cut and gradual transition. At the early phase of this study, researchers pay much attention on the detection of hard cut, while with the development of video technology, more focus has been drawn on the gradual transition detection and recognition.This paper summarizes a large number of advanced theories in this field, based on which we propose a new algorithm with the combination of adaptive threshold and fitting feature. The new algorithm not only can detect hard cut effectively, but also do well in detecting the boundary of gradual transition and recognizing the type of gradual transition. The first thing of this algorithm is to extract the feature description of video. Due to the massive size of video, it is really time-consuming and inefficient to detect boundary on the original data. In order to make the process of detection and verification more convenient, we need a light-weighted representation of video. Frame is divided into 3 × 3 sub-block according to the ratio λ:β:λ. Color histogram is then accumulated on the sub-block, and every frame will have 9 histograms, we call them the histogram group. Similarity between frames is then calculated based on the weighted distance of histogram independently. The feature description is achieved when the processing of the video has been finished.This paper proposed adaptive threshold to avoid using the threshold set manually. Adaptive threshold is derived after certain process on the feature description of original video. Once we exclude the interference of outside, the performance of adaptive threshold becomes robust and excellent.Before the detection of gradual transition, we picked up some standard gradual transitions carefully, and using Fourier Fitting to extract their fitting features. Finally, gradual transition templates were made after the normalization composed on these fitting features. With the help of reverse-order algorithm, we can detect the candidate boundary, and extract their fitting features to compare with templates. Once matched, the candidate is considered to be a confirmed boundary, and has the same transition type with the template, otherwise, the candidate is taken as invalid boundary.In order to accelerate the extraction of feature description, CUDA-GPU is deployed to accumulate histogram parallel. Only 20ms is need to process a frame with the resolution of 1920× 1080, and the performance is improved vastly.