窗口融合特征对比度的光学遥感目标检测

提出了一种基于窗口融合特征对比度的光学遥感目标检测方法。首先,在训练图像上生成大量不同尺寸的滑动窗,计算了各窗口的多尺度显著度、仿射协变区域对比度、边缘密度对比度以及超像素完整度4项特征分值,在确认集上基于窗口重合度和后验概率最大化学习各个特征的阈值参数。然后,采用Naive Bayes框架进行特征融合,并训练分类器。在目标检测阶段首先计算测试图像中各窗口的多尺度显著度分值,初步筛选出显著度高且符合待检测目标尺寸比例的部分窗口。然后计算初选窗口集的其余3项特征,再根据训练好的分类模型计算各个窗口的后验概率。最后,挑选出局部高分值的候选区域并进行判断合并,得到最终目标检测结果。针对飞机、油罐、舰船等3类遥感目标的检测结果显示:4类特征在单独描述3类目标时表现出的性能各有差异,最高检测准确率为74.21%~80.32%,而融合方案能够综合考虑目标自身特点,准确率提高至80.78~87.30%。与固定数量滑动窗方法相比,准确率从约80%提高到约85%,虚警率从20%左右降低为3%左右。最终高分值区域数降低约90%,测试时间减少约25%。得到的结果显示该方法大大提高了目标检测精度和算法效率。

Optical remote sensing object detection based on fused feature contrast of subwindows

A detection algorithm for optical remote sensing targets was proposed based on the fused features contrast of subwindows. Firstly, a large number of varisized sliding windows were generated in a training image, and four types of scores related to multi-scale saliency, affine invariant region contrast, edge density and superpixel straddling were computed within each window. The feature parameters were learned on validation sets by maximizing localization accuracy and posterior probability. Then, all the features were combined in a Naive Bayesian framework and a classifier was trained. In the target detection step, the multi-scale saliency score was firstly computed within all the windows of test images, and partial windows with higher saliency and proper sizes matching to the objects to be detected were selected preliminarily. Furthermore, other scores were computed within the selected windows, and the posterior probability of each window was computed by using the trained classifier. Finally, windows with high local scores were selected and merged and the final detection results were obtained. The detection experiments were performed on three types of remote targets including planes, oilcans and ships, and the results show that each type of feature appears different properties for targets described, the highest accuracy is 74.21% to 80.32%. The proposed method outperforms all the single feature methods and the accuracy is improved to 80.87% to 87.30%. By compared with the fixed number sliding window algorithm, the accuracy rate is improved from about 80% to 85% and the false alarm rate is reduced from about 20% to 3%. Furthermore, the proposed method shows a 90% reduction in the number of windows and 25% reduction in the detection time due to the selection in the intermediary stage. It concludes that the method improves detection accuracy and algorithm efficiency greatly.