基于自适应多尺度图卷积网络的多标签图像识别

利用一阶谱图卷积探索类别标签间关系是目前多标签图像识别常用的手段,但是,较多的图卷积层数易出现过度平滑现象,使得该方法存在局限性.为此,提出一种基于自适应多尺度图卷积网络的多标签图像识别方法,主要思路为:采用块Krylov子空间形式的谱图卷积来挖掘类别标签间的相关性,在每个图卷积层中拼接多尺度信息并扩展到深层结构,并在自适应标签关系图模块所构建的关系图上学习分类器,从而更加有效地进行多标签图像识别.通过两个公开数据集PASCAL VOC 2007和MS-COCO 2014上的实验结果验证了所提出方法的有效性.

Multi-label image recognition based on adaptive multi-scale graph convolutional network

Utilizing the first-order spectral graph convolution to explore the correlation between category labels is a common method for multi-label image recognition. However, more graph convolution layers are prone to over-smoothing, which causes some limitations for the method. With respect to the aforementioned problem, a multi-label image recognition method based on the adaptive multi-scale graph convolutional network is proposed. The main idea is as follows: the spectral graph convolution in the form of block Krylov subspace is employed to mine the correlation between category labels, and the multi-scale information eixsted in the convolutional layer is spliced and extended to the deep structure. At the same time, the classifier is learned on the relation graph constructed by the adaptive label relation graph module, accordingly the multi-label image recognition is performed more effectively. Experimental results on two public datasets including PASCAL VOC 2007 and MS-COCO 2014 verify the effectiveness of the proposed method.