Research On Methods For Caption Text Extraction In Video Retrieving

With fast growth of video data, how to index and retrieve large-volume video data has become a popular subject. There is huge amount of text information contained in video contents, which serves to describe, explain and supplement video image and thus develops into an important clue for video retrieving and indexing. Text-based video indexing and retrieving could be realized by establishing a link between text and video contents, which involves text information extraction.Text information extraction is composed of text localization and text segmentation. In this work, study is conducted on a hybrid-feature-based method for localizing caption text and detailed explanation is given. Among existent hybrid-feature based methods, some simultaneously detect multiple features and merge the result, while others detect each feature in order of its efficiency, but all the results are not satisfied. The composition of proposed scheme is as follows: regarding its high recall rate, edge detector is used to obtain candidate text blocks; then connected-component analysis is conducted to correct the bounding boxes of text block and finally Support Vector Machine (or SVM for short) is utilized to filter out non-text blocks.Considering time and large-size memory cost during the SVM training, the proposed method reduces the training samples before the training. Based on the existent methods, the method selects a portion of samples from the whole training samples depending on the Euclidean distance between samples from different categories and also the same category. According to the experimental result, the proposed method can achieve ideal training result with reduced samples.As for text segmentation, we provide a method to build a joint probability model using color and texture feature of the character. The existent methods for selecting samples to build model are mainly based on the edge feature of the character, which is not precise enough. So we improve this by utilizing the Chinese character's stroke feature due to the fact that the width of stroke is almost same for each character with the same height. Experimental results show that the model we build can segment text pixel from background pixel successfully.