| Summary: | 碩士 === 國立臺灣科技大學 === 資訊工程系 === 96 === Recently, Chien et al. presented an efficient predictive watershed-based video segmentation algorithm. This thesis presents an improved predictive watershed method for segmenting objects. Due to inheriting more segmentation results from the current frame, our proposed improved method needs less computation e®ort when compared to Chien et al.'s method. In addition, an e±cient region merging strategy is employed to alleviate the over-segmentation problem. Finally, a new efficient automatic renewal scheme, which is based on the integrated error of region-pairs accumulation error criterion, is presented to interrupt the error propagation. Experimental results demonstrate the execution time and quality advantages of our proposed improved algorithm. Under five video sequences, the execution-time improvement ratios of our proposed algorithm over the traditional watershed-based approach and the the previous algorithm by Chien et al. are 77.4% and 41.8%, respectively, in average.
|
|---|
