Abstract: Web images generally consist of heterogeneous information including texts, colors and textures. This paper proposes a new method, called local co-training (LCT), for semi-supervised classification of web images based on fusion of heterogeneous information. The proposed method employs a set of local linear models to represent data points of each view, and incrementally refines these models by exploiting unlabeled data with information propagation and co-training. The local co-training builds a bridge between graph-based methods and co-training. The local co-training can model the instance distribution accurately in the high-dimensional space, and learn local models incrementally, which benefits the online classification of large scale of web images. Experiments on Corel, Pascal and ImageNet datasets demonstrate that the local co-training can effectively improve the classification performance of learners by exploiting multiple attribute sets and unlabeled data.