返回Google圖書搜尋

A Cognitive Vision Approach to Image Segmentation

出版	INTECH Open Access Publisher, 2008
ISBN	95376190369789537619039
URL	http://books.google.com.hk/books?id=94LgoAEACAAJ&hl=&source=gbs_api

註釋In this chapter, we address the problem of image segmentation with a cognitive vision approach. More precisely, we study three major issues of the segmentation task in vision systems: context adaptation, selection of an algorithm and tuning of its free parameters, according to the image content and to the application needs. Most of the time, this tedious and time-consuming task is achieved by an expert in image processing using a manual trialand-error process. Recently, some attempts at automating the extraction of optimal parameters of segmentation have been made but they are still too application-dependent. The re-usability of such methods is still an open problem. We have chosen to handle this issue with a cognitive vision approach. Cognitive vision is a recent research field which proposes to enrich computer vision systems with cognitive capabilities, e.g., to reason from a priori knowledge, to learn from perceptual information, or to adapt its strategy to different problems. We propose a supervised learning-based methodology for off-line configuration and on-line adaptation of the segmentation task in vision systems. The off-line configuration stage requires minimal knowledge to learn the optimal selection and tuning of segmentation algorithms. In an on-line stage, the learnt segmentation knowledge is used to perform an adaptive segmentation of images. This cognitive vision approach to image segmentation is thus a contribution for the research in cognitive vision. Indeed, it enables robustness, adaptation, and re-usability faculties to be fulfilled. Finally, by addressing the problem of adaptive image segmentation, we have also addressed underlying problems, such as feature extraction and selection, and segmentation evaluation and mapping between low-level and high-level knowledge. Each of these well-known challenging problems is not easily tractable and still demands to be intensively considered. We have designed our approach (and our software) to be modular and upgradeable so as to take advantage of new progresses in these topics. The brittleness of our approach to unknown situations is currently its major drawback. This concerns the context analysis level as well as the segmentation level. The concerned algorithms are the DBScan algorithm for image-content clustering and the SVMs for the semantic segmentation. Currently, neither the clustering algorithm nor the SVMs are able to adapt dynamically to new training data: the learning process must be run again on the whole training data set. The use of incremental machine learning techniques should be useful to fulfil the property of continuous learning. The main idea of incremental learning.