基于聚类分析和集成神经网络的序列图像多目标识别算法

Multi-objects Recognition Algorithm in Sequence Images Based on Clustering Analysis and Ensemble Neural Networks

针对现有的集成神经网络的训练子集选择时没有考虑样本空间的分布情况,使得构造的训练子集具有很大的随机性和主观性,集成的差异性不能有效保证的缺点,提出了一种新的基于Hu七阶矩、RPCL聚类分析和集成神经网络的序列图像多目标识别算法。该方法首先在训练视频中连续提取序列图像中的目标——人、人群、汽车,利用Hu七阶矩提取轮廓信息。为了防止Hu七阶矩对小目标和非刚体目标的描述能力弱的缺点,再提取图像的面积信息。其次对所提取的8维数据采用基于对手惩罚策略的竞争学习算法(RPCL)进行聚类分析,得到待分样本的分布。再次采用提出的单个神经网络生成算法得到单个神经网络。最后采用相对多数方法对神经网络进行集成。采用基于boosting,bagging方法的集成神经网络和该算法进行比较,结果表明该方法的分类精度要高于传统方法,是一种有效的目标识别算法。

These ensemble neural networks have great randomicity and arbitrary when choosing samples to build up the training subsets since they don＇t consider the distribution information of theses samples. The performance of ensemble will deteriorate greatly once the training subsets are not appropriately selected. So in practical application, the training subsets must include many instances, which result in great needs of computation time and storage space. In order to overcome above drawbacks, this paper put forwards a new ensemble neural networks approach based on RPCL clustering analysis and Hu invariable moment. Firstly, we extracted targets from sequence images, such as pedestrian, crowd, car from single frame image in training video. The outline information was extracted by Hu invariable moment, in order to avoid the default that Hu invariable moment can not describe small targets and non-rigid targets, area information are extracted too; secondly,8 dimension data about these samples were clustered using RPCL algorithm and obtain their distribution; thirdly, single neural network was constructed through this method proposed by this paper; finally, en- semble was constructed by relative majority voting method. We adopted ensemble neural network based on boosting and bagging method to compare this method,experiment result illustrate that the method has a high classification precision comparing with tradition method, and it is a effective target recognition approach.