融入极值点特征的深度交互式图像分割方法研究

Research on Deep Interactive image Segmentation Method Incorporating Extreme Points Feature

交互式图像分割是指向计算机提供有用的先验知识,通过用户交互辅助计算机把感兴趣的区域从复杂环境中分离出来。交互式图像分割目前存在的两个难点:一是用户交互过程复杂,操作不方便。二是计算机根据用户提供的交互信息分割出的结果不理想。针对上述问题,提出了一种融入极值点特征的深度交互式图像分割方法。首先通过用户标出图像中目标区域顶部、底部、最左侧、和最右侧的极值点,然后利用算法求解出以极值点为顶点的极值框;根据欧氏距离变换原理,将极值框求解成欧式距离映射图,最后将欧氏距离映射图和图像的RGB三个通道级联输入到卷积神经网络(Convolutional Neural Networks,CNN),通过卷积神经网络提取特征,输出特征图。与其他类似的方法相比较,该方法用户交互时间少、分割结果更加完整。

Interactive image segmentation is to provide useful prior knowledge to the computer,and to separate the region of interest from the complex environment by user interaction assistant computer.At present,there are two difficulties in interactive image segmentation:Firstly,the user interaction is time-consuming and the segmentation efficiency is low.Secondly,the results of computer segmentation based on the interactive information provided by users are not ideal and the segmentation accuracy is low.Aiming at the above problems,this paper proposes a depth interactive image segmentation method which incorporates the feature of extreme points.First,the user marks the extreme points on left-most,right-most,top,bottom points of the target area in the image,and then the algorithm is used to solve the extreme box with the extreme points as the vertex.According to the Euclidean distance transformation principle,the extreme box is transformed into a Euclidean distance map.Finally,the Euclidean distance map and the RGB channels of the image are cascaded into a Convolutional Neural network(CNN)to extract feature and output feature map through the Convolutional Neural network.Compared with other similar methods,this method has less user interaction time and more complete segmentation results.