Efficient object segmentation for images and video sequences

Abstract

Most computer vision applications require segmentation of objects or regions from an original image in order to understand/analyze a given image frame. In a variety of applications, such as surveillance systems, fire control systems, guidance, robotics, and autonomous vehicle navigation, object segmentation (region partitioning) is necessary to provide basic information for high level image analysis, retrieval, and recognition systems. Segmentation performance is usually determined by the accuracy and time required for region partitioning. Accuracy can be determined by assignment of an appropriate color for a region of the original image. The time required can be determined by the processing time necessary for a final segmentation result. For real time applications using vision sensors, fast techniques for segmentation are necessary. An efficient object segmentation algorithm is proposed to improve the accuracy of segmentation and reduce the time required. For faster region segmentation, a fast image segmentation scheme is pro-posed based on multi-resolution analysis (MRA) and wavelets. Many spatial segmentation algorithms use a pre-selected feature, such as color/intensity or edge/direction. These features are transformed into a probabilistic distribution. Usually, the transformed distribution of the given feature provides information that is sufficient for segmentation of image regions. An efficient algorithm for image segmentation based on a multi-resolution application of a wavelets transform is proposed in which the original feature space is transformed into a lower resolution image with a wavelets transform in order to derive fast computation of the optimum threshold value in the feature space. A single feature value or multiple feature values are determined as the optimum threshold values based on this lower resolution version of the given feature space. The optimum feature values in the lower resolution image are projected onto the original feature spac...