基于隐式模型表示的对称物体检测算法

摘要

This paper proposes a novel image rotation object detection algorithm,which is especially suitable to detect objects having non-regular and random rotation symmetry characteristics in real-world images.Based on the implicit model representation,the algorithm counts the spatial distribution of the key points in the image and estimate the rotation center of the object.The algorithm extracts the visual interesting points from the image middle-level features.The unsupervised learning is employed in the key point feature space,aiming at labeling each position of the key points as one of the symmetry feature clusters.Thereby the probability map for center is gained by summing all the values voted in every cluster under some given radius.The maps are weighted summed together to obtain the global rotation center probability map,and the coordinates of the rotation center are achieved by saliency detection method on the map. Experimental results show that the algorithm can effectively detect the irregular rotational symmetry objects in real-world images,and it also has a higher accuracy for the estimation of the rotational symmetry center.%针对自然界中具有旋转对称特征物体的不规则性和随机性特点，提出一种新的图像旋转目标检测算法。使用基于隐式模型表示的方法来统计图像中关键点的空间位置分布，并在此基础上对物体的旋转中心进行估计。通过中层特征对图像中视觉感兴趣的关键点进行提取，在所提取关键点的特征空间进行无监督学习，将这些关键点所在空间位置聚类为不同的旋转特征簇，进而通过各个特征簇的投票值，获得在特定半径下的中心概率映射图。使用多半径尺度的显著度检测对各个半径下得到的概率映射图进行加权叠加，获得图像全局的旋转中心概率映射图，并通过显著度检测算法获取旋转中心坐标。实验结果表明，该算法对自然图像中不规则的旋转对称物体都能够进行有效的检测，对旋转对称中心的估计也具有较高的精度。