基于自适应流形相似性的图像显著性区域提取算法

An image saliency object detection algorithm based on adaptive manifold similarity

为了在图像显著性区域提取过程中改善算法的自适应性和精准度,提出基于自适应流形相似性的图像显著性区域检测算法。将图像分割成超像素,根据图像中显著性区域频率变化比较大的特性,生成图像显著性区域的高频节点;针对高频节点利用凸包运算寻找显著性区域的种子节点;使用流形算法在图像中对种子节点进行显著性区域信息扩散,得到图像的显著性区域。试验结果表明:利用流形算法搭建求解每个数据的邻接矩阵进行信息扩散,能够在保证信息精准分类的同时提高算法的自适应性,其结果优于同类的图像显著性区域检测算法。

In order to improve the adaptability and precision in extracting salient regions in images, an image salient re- gion detection algorithm was proposed based on adaptive manifold similarity. An input image was segmented into super- pixels which were represented as the nodes in a graph. The node with high frequency was generated by the characteris- tics of the salient regions. Convex hull computation was used to generate the saliency seeds of the salient object area ac- cording to high-frequency nodes. The proposed algorithm was used to complete information reconstruction of the current image by adaptively assessing the salient weights on the edges between the nodes. In addition, based on local character- istics information reconstruction, the proposed algorithm utilized similarity extraction function to self-adaptively obtain the similarity characteristics and manifold structures in order to spread salient characteristics information. The experi- mental results showed that the quadratic programming solution exploited to compute the weights between the nodes could effectively avoid threshold selection and enhance robustness accordingly, and the proposed method performed better than the other state-of-the-art methods.