Text detection and localisation in video

Abstract

In this dissertation a caption detection and localisation algorithm for News videos is proposed. The proposed method is divided into several steps and one important feature of this algorithm is that it does not process all the video frames in order to extract the caption. This makes the algorithm very efficient when compared to other text detection algorithms. The proposed method processes only four frames every two seconds and uses the logical AND operation to eliminate non-static and redundant objects. It then uses a morphological operator to generate text blobs. The proposed algorithm managed to achieve a Recall (R) rate of 79.92%, a Precision (P) rate of 79.35% and Quality (Q) of bounding box rate of 92.10%. The result obtained for the quality of bounding box is very high and this is important since it will aid in the post processing stages when binarization is performed on the caption region. Both negative and positive text polarities can be detected and the colour of the captions does not affect the results obtained. This algorithm’s main objective is to detect captions found in the TVM news bulletin however it can be also used for other news videos. TVM is a local Maltese TV channel which amongst many shows, transmits News bulletins.