融合神经网络与超像素的候选区域优化算法

为解决目标检测中候选区域召回率低的问题,提出融合神经网络与超像素的目标候选区域算法。该算法利用神经网络提取更能清楚表达目标边界的特征,并使用聚类、相似性等策略,计算每个滑动窗口所含有的边缘信息量;将待测图像使用简单线性迭代聚类算法分割成若干个超像素,并利用超像素的空间位置、完整性、相邻超像素间的对比度信息,计算各个超像素的显著性得分及每个滑动窗口的显著性得分;根据每个滑动窗口的边缘信息及显著性得分筛选滑动窗口。在PASCAL VOC 2007测试集上进行对比实验,其实验结果表明:所述算法能够快速产生定位质量高的候选区域。

Region proposals optimization algorithm combining neural networks and superpixels

In order to solve the low recall problem of the region proposals in object detection, the object region proposals algorithm, which combines neural networks and superpixels, was proposed. The edge features, which can be represented clearly by neural networks, were extracted from the images to be detected, and the score of edge information for per sliding window was computed by the strategy of edge clustering and the affinities between the edge groups. The several superpixels of this images were obtained by simple linear iterative clustering algorithm, and the salient object score of a superpixel was calculated using the location, integrity of this superpixel and the contrast with neighbors. The salient objects score of per sliding window was received by these saliency scores of superpixels according to the Euler distance strategy between the sliding window and these superpixels. The region proposals were determined by two components including edge information scores and salient object scores. The comparative experiments were conducted in PASCAL VOC 2007 test set, and the experiment results show that the proposed algorithm can fast generate a set of region proposal with higher localization.