改进的多尺度图谱和局部谱的目标提取算法

Object extraction algorithm based on improved multi-scale spectral and local spectral

当目标对象与背景的纹理较多或两者纹理较接近时,基于多尺度图谱和局部谱的目标提取算法不能很好地提取目标,主要由于在计算相似度度量时,金字塔多尺度图谱算法特征选取较简单。针对算法不足,提出基于改进的金字塔多尺度图谱和局部谱相结合的目标提取算法,主要通过改进多尺度图谱中干涉轮廓权重构造方法,原算法中是基于拉普拉斯边缘图和梯度图,改进后是基于多尺度边缘概率检测算子和方向分水岭算法产生的边缘强度图。多尺度边缘概率检测算子可以有效地解决纹理较复杂图像分割不佳问题,方向分水岭算法可以有效解决由于目标和背景部分边界信息较接近导致的分割不佳问题。实验结果表明改进算法有效地弥补了原算法的不足,并且具有良好的目标提取效果。

When textures of object and background are more or their textures are similar, object extraction algorithm based on multi-scale spectral and local spectral can’t be very well, the main reason is that feature selection of pyramid multi-scale spectral algorithm is relatively simple when calculating the similarity measure. Aiming at drawback of this scheme, the new scheme based on improved pyramid multi-scale spectral and local spectral is proposed. The main area of improve-ment is weight construction method based on intervening contours. The original scheme is based on the Laplace edge map and the gradient map, while the improved scheme is based on the edge strength map which is generated by multi-scale probability of a boundary detection operator（mPb operator）and Oriented Watershed Transform algorithm（OWT）. The mPb operator can effectively solve the problem that the result of segmentation about complicated texture image is not good. Oriented watershed transform algorithm can effectively solve the poor segmentation problem when part boundary information of target and background are similar. Experimental results demonstrate that the new scheme is superior com-pared to the original scheme in the aspect of object extraction of effect.