Study Of Methods For Video Caption Detection And Extraction

A key technology of video caption detection and extraction for video indexing and retrieval is studied in this paper, and three algorithms are proposed: (1) A method for video caption extraction based on adaptive location and FCM segmentation: First video shot and text event are detected, and the caption region is located by the adaptive and projective method. Frame average method is employed in a series of video frames with the same caption for enhancing character. Then, the obtained colorful caption region is segmented with Fuzzy c-means method. (2) Automatic video caption detection and recognition based on spatial-temporal information: The caption is detected using quantized spatial difference density (QSDD). Then the difference image between adjacent frame pairs is calculated at each caption transition, and caption region is located through the extracted features from difference image using the fuzzy clustering neural network (FCNN). Caption enhancement can be employed to reduce the influence of the complex background. Finally the caption region is separated into individual characters, and each of which is filtered to eliminate the remaining background and is enlarged by interpolation method. (3)Video caption extraction method based on temporal feature vectors: a binary abstract sequence from a video segment is generated using temporal feature vectors. By the analysis of the pixel changes in the adjacent frame pairs, the changing frames with captions can be effectively located. Then a temporal feature vector for the frames with the same caption is constructed in the video segment, and the caption can be extracted from background pixels using the temporal feature vector. The experiment results demonstrate the superiority of the proposed detection and extraction scheme in both the accuracy and robustness. In addition to the location of caption region is the key process of many algorithms on video caption extraction, the theory and whole processes of three methods are presented in this paper. (1) The method based on region connectivity analysis and matched template for locating of news video caption. (2) Video caption location based on wavelet transform. (3) Automatic video caption location based on support vector machine.