为满足自然场景下显著性检测精度的要求, 提出了一种显著性检测优化方法。该方法采用简单线性迭代聚类分割算法将图像分割为多个超像素区域, 并提取颜色区域对比度特征。通过Harris角点检测算法定位目标的大致几何中心, 以中心概率的形式表征目标空间分布特征, 并进行目标位置自适应的特征融合。基于目标空间分布特征和图像灰度重心, 实现抑制背景、增强目标的显著图优化; 利用针对显著性值的空间平滑技术, 可增加显著图的连续性。实验结果表明, 该方法在几个公开的测试集中的测试具有较高的准确率、召回率和较低的平均绝对误差, 可应用于复杂自然场景下的显著性检测。

A new saliency detection optimization method is proposed to satisfy the accuracy requirement of saliency detection in the natural scene. The method can divide an image into multiple superpixel areas using the simple linear iterative clustering algorithm, and extract the contrast feature of color regions. The general target geometric center is located by the Harris corner detection algorithm. The center probability is used to describe the target space distribution feature, and the adaptive feature fusion for the target location is carried out. Optimization of a saliency map with background suppression and target enhancement is realized based on target space distribution feature and image gray centroid. The continuity of the saliency map can be enhanced by the space smoothing technique for the saliency value. Experimental results show that the test with this method does not only have high precision rate and recall rate, but also has low mean absolute error in several testing sets, and the method can be applied to the saliency detection in complex natural scenes.