基于主动轮廓模型的自然图像分割研究

The research purpose and significance of this project lies in: first, the selective attention mechanism of human brain is further combined with active contour model to enrich and develop the theory system of active contour model; Second, the texture feature descriptions of natural images is substituted for intensity homogeneity which is emphasized by the traditional active contour model methods. It will broaden the application scope of the active contour model in the natural image segmentation; Third, the fast and effective framework of active contour model with automatic initialization is designed for natural image segmentation from the technical level. By this way, we hope to promote the application of the active contour model in the field of image processing, computer vision science, etc. The expected achievements of this research project are: (1) To present the saliency detection algorithm based on the patches in subsapce; (2)To analyze the advantages and disadvantages of our active contour model methods based on the selective attention mechanism of human brain in depth, we combine the clustering idea to design the automatic contour initialization method based on the edge information of the salient object for two advantages: first, to reduce the interference of background information to the contour evolution and the convergence time ;(3) To make the contour model more robust and applicable to effective segmentation of natural scene images in complex environments, this project will develop a texture representation method of the natural images and a twice kernal sparse representation algorithm representing the texture features, simultaneously.

进行该项目的研究目的和意义在于：第一、进一步将人脑选择性注意机制与主动轮廓模型研究结合起来，丰富和发展主动轮廓模型的理论体系；第二、将自然图像的纹理特征替代传统主动轮廓模型方法强调的灰度一致性的性质，拓宽主动轮廓模型在自然图像分割的应用范围；第三、从技术层面上设计快速高效的适用于自然图像目标分割的自动初始化主动轮廓模型框架，推动主动轮廓模型在图像处理、计算机视觉等多学科领域的应用发展。本项目研究的预期成果是：(1)提出基于子空间的分块显著性提取方法；(2)深入分析本课题组提出的基于人脑选择性注意机制的主动轮廓模型方法的优缺点，结合聚类思想设计基于显著目标边界信息的自动初始化轮廓方法，降低复杂背景对轮廓演化的影响和轮廓迭代收敛时间；(3)为了使主动轮廓模型更具鲁棒性且更适应于复杂环境下自然场景的有效分割，提出自然图像的纹理表示并基于二次核稀疏描述的纹理特征描述。

自然图像具有数据量大、结构复杂性、自然图像目标外观的变化大等特点，导致基于主动轮廓模型的图像分割方法针对复杂结构的自然图像存在较大局限性。因此，本项目将人脑选择性注意机制与主动轮廓模型研究结合起来，丰富和发展主动轮廓模型的理论体系，设计快速高效的适用于自然图像目标分割的自动初始化主动轮廓模型框架；提取自然图像目标的显著边界点，并提出抠像算法对边界点进行精细化凸包操作，得到主动轮廓模型的初始轮廓；定义显著性提取为在图中标记的重构任务，提出一种基于线性邻域重构的显著性提取方法；将火焰这一特定目标检测视为自底向上和自顶向下相结合的视觉选择注意的过程，提出一种基于视觉显著性的视频火焰检测方法；结合了GAC模型和LRCV模型的优点，提出了基于新的符号压力函数的边界局部区域型模型并提出一种核化的局部PC模型。本项目提出的算法，在自然图像目标分割和显著目标检测方面的查准率、查全率和F-measure值上都优于其它主流算法。上述研究成果使得主动轮廓模型更具鲁棒性且更适应于复杂环境下自然场景的有效分割，推动主动轮廓模型在图像处理、计算机视觉等多学科领域的应用发展具有重要意义。