一种用于图像分类的改进的有偏特征采样方法

An improved biased sampling strategy for object categorization

为了模拟图像分类任务中待分类目标的可能分布,使特征采样点尽可能集中于目标区域,基于Yang的有偏采样算法提出了一种改进的有偏采样算法。原算法将目标基于区域特征出现的概率和显著图结合起来,计算用于特征采样的概率分布图,使用硬编码方式对区域特征进行编码,导致量化误差较大。改进的算法使用局部约束性编码代替硬编码,并且使用更为精确的后验概率计算方式以及空间金字塔框架,改善了算法性能。在PASCAL VOC 2007和2010两个数据集上进行实验,平均精度比随机选取的特征采样方法能够提高约0.5%,验证了算法的有效性。

In order to simulate the probability distribution of the objects in image classification task and make the feature points focus on the object region, an improved version of Yang＇ s biased sampling strategy is proposed. The original algorithm combines the probability map of the object occurrence and saliency map to compute a probability distribution map for feature sampling, uses the hard voting VQ method for the region features＇ coding, which leads to large quantization error. Our algorithm uses locality-constrained linear coding and a better posterior probability computing method instead, and moreover, uses the SPM framework to improve algorithm performance. Experiments on PASCAL VOC 2007 and 2010 dataset show the algorithm has better performance than random sampling method.