Caption Localization and Detection for News Videos Using Frequency Analysis and Wavelet Features

In this paper, we propose an algorithm to detect captions from news videos. The propose method only detects captions excluding other miscellaneous types of text. The algorithm makes use of the fact that the text remains in many consecutive frames to reduce the number of the processing frames. The caption beginning frame is detected first, then a caption candidate strip in the caption beginning frame is defined. Moreover, the difference of the caption candidate strip between consecutive frames is computed, and then the difference information is transformed to frequency domain by discrete cosine transform. Frequency analysis is used to define the caption candidate region, and twelve wavelet features are extracted from the region and considered as the input of the classifier to detect the text blocks. Experimental results show that the proposed approach can fast and robustly detect captions from news video.