有限样本下基于图卷积神经网络的目标检测方法研究

Object detection method based on graph convolution net under limited samples

当前目标检测算法的成功,很大程度取决于大量数据的支撑。当数据量较少时,多数目标检测算法无法达到较好的性能水平。有限样本下的目标检测主要研究有标签样本数量较少情况下的目标检测算法。提出一种新的有限样本下的目标检测方法,在具有大量标注信息的基类数据集上训练得到一个具有较强语义特征提取能力的特征提取器以及候选区域推荐网络,将该网络应用于只有少量标注样本的新类数据上训练候选区域推荐网络,同时将所得到的候选区域特征构建成一个目标图,通过采用图卷积神经网络在模拟有限样本的场景下进行训练,从而获得新类测试目标的候选框偏移量以及目标类别。实验证明,所提方法能够解决有限样本目标的分类和定位问题,比采用通用目标检测方法具有更好的泛化能力,具有广泛的潜在应用。

The success of current object detection algorithms largely depends on the large amounts of data.When there is limit amount of available data,most object detection algorithms cannot achieve good performance.The task of object detection under limited samples mainly studies object detection algorithms based on a few numbers of label samples.This paper proposes a new object detection method for limited annotated samples.The core of the method is to build an object graph based on the predicted bounding box of the potential target objects in an image.By using the graph convolutional neural network,the category is predicted by learning the similarity of object features from the limited annotated information.Through training on the base class data set with a large amount of labeled information,a feature extractor with strong semantic feature extraction capabilities and a candidate region recommendation network are obtained,and the network is applied to train candidates on new class data with only a few labeled samples.The regional recommendation network,at the same time,constructs a object graph based on the obtained feature of the candidate area.Through training of the object graph by using the graph convolutional neural network,the candidate frame offset and target category of the new test target is learned.Experiments have demonstrated that this method can solve the problem of classification and localization of limited sample targets,and has better generalization ability than the object detection based method.