基于多尺度视觉特征组合的高分遥感影像目标检测

Target Detection from High Resolution Remote Sensing Images Based on Combination of Multi-scale Visual Features

针对高空间分辨率(高分)遥感影像中目标空间结构明显、颜色丰富的特点,提出一种基于多尺度视觉特征组合的目标检测方法.首先采用两组特征描述子来表达目标的颜色与空间结构特征,其中,一组是彩色变换后的密集空间金字塔尺度不变特征变换(scale-invariant feature transform,SIFT)描述子,另一组是多尺度方向梯度直方图(histogram of oriented gradient,HOG)与色度、饱和度、明度(hue,saturation and value,HSV)颜色模型组合的特征描述子,然后用支持向量机(support vector machine,SVM)对两组特征分别进行模型训练与目标检测,对于彩色变换密集空间金字塔SIFT描述子,采用视觉词袋空间金字塔匹配核(spatial pyramid matching kernel,SPMK)构建SVM的输入,对于HOG与HSV特征,直接将特征向量按顺序合成作为SVM的输入,最后取两组检测结果的交集作为最终检测结果.为验证方法的有效性,分别以航空影像中的建筑物和鱼排,以及高分卫星影像中的游艇为目标进行检测.结果表明,该方法的目标检测精度可达到90%以上.

To utilize the spatial structure and color information of the target in the high resolution remote sensing images, this paper proposes a method of target detection based on multi-scale visual feature combination. Firstly, we use two types of feature descriptor to express the color and spatial structure information of targets, one is the dense spatial pyramid scale-invariant feature transform （SIFT） descriptor in spatial pyramid of the color transformed images, the other is the multi-scale histogram of oriented gradient （HOG） feature combined with the hue, saturability and value （HSV） features. Secondly, the two sets of features are used to train classifier through support vector machine （SVM）. For the color transformed dense SIFT spatial pyramid feature descriptor, the input of SVM is constructed by spatial pyramid matching kernel （SPMK）. For the combined HOG and HSV feature, the orderly-combined eigenvectors are taken as the input of the SVM classifier.Finally the intersection of the two classification results are regarded as the final results.For verification, we tested the buildings and fish rafts in aerial images and the yachts in satellite images. Experimental results show that the proposed method can improve the precision of target detection, which can reach to by 90% or better.