A Real-time Semi-automatic Video Segmentation System Based on Mathematical Morphology

摘要

Mathematic morphology provides a systematic approach to analyze the geometric characteristics of signals or images, and has been widely applied to many applications such as edge detection, object segmentation, noise suppression. In this paper, a supervised morphology based video segmentation system is proposed. To find where a semantic object resides, the user could click the mouse near the boundary of the object in the first frame of a video to indicate its rough definition, shape and location. The proposed system will automatically segment the first frame by first locating the searching area and then classifying the units in it into object part and non-object part to find out the continuous contour by means of a multi-valued watershed algorithm using a hierarchical queue. An adaptive morphological operator based on edge strength, which is computed by a multi-scale morphological gradient algorithm, is proposed to lower the error of user assistance such that the searching area is created correctly. Once extended to video object segmentation, a fast video tracking technique is applied. Under the assumption of small motion, the object can be segmented in real-time. Moreover, an accuracy evaluation mechanism is proposed to ensure the robustness of the segmentation.