基于视觉显著性及多特征分析的目标检测

Target Detection Based on Visual Saliency and Multi-feature Analysis

针对视觉显著性分析不能辨别目标且单个特征描述目标具有局限性的问题,提出基于视觉显著性及多特征分析的目标检测.首先,对已标定训练图,生成遍历整幅图像的随机采样区域,通过多特征分析获取每个区域包含目标可能性的先验参数信息;然后,对测试图,依据上述先验信息,基于贝叶斯模型计算每个随机采样区域包含目标可能性的评分值,并将值高的若干区域标记为目标候选区域;最后,结合显著性分析及判别准则,对候选区域进一步判定,以确定最大可能涵盖目标的区域,从而实现目标检测.研究结果表明:显著性分析具有对目标所在区域的主动选择性;多特征结合能有效描述目标以使目标更具可区分性.

Due to the difficulty to indicate an object for the existent visual saliency models and the limitation of representation ability with single feature,an algorithm of object extraction was projected combined with multi-feature and visual saliency analysis.Firstly,for the train images,the random windows uniformly distributed over the entire image were sampled and the prior information of parameters were learned through the analysis of multi-feature.Secondly,for the test image,the random windows were sampled too and their scores were calculated integrated the analysis of characteristics above in Bayes Model.Finally,the regions output by this method were arranged according to their weights and the region most possibly contained object was determined combined with the analysis of saliency and its criterion,thus the target detection was achieved ultimately.The results showed that the algorithm has an active selectivity to the candidate regions depended on the analysis of visual saliency and has a distinguish ability between target and non-target due to the combination of multi-feature.