SPNet: Superpixel Pyramid Network for Scene Parsing

Scene parsing is the important part of computer vision research. And the deep coding-decoding network is widely applied to scene parsing. However, there are still some problems, such as ambiguity of object edge segmentation and uncertainty when segmenting small-size-objects in scene analysis. In this paper, we propose Superpixel Pyramid Network for Scene Parsing. First, a deep coding-decoding network is used to learn image features. Then, multi-scale spatial pyramid pooling structure is employed to enhance the performance of small-size-objects. Next, the Superpixel Segmentation is also applied to cope with the problem of ambiguity of object edge. Finally, a two-layer neural network classifier is applied to identify the fused features pixel-by-pixel. Extensive experimental results over ADE20K, PASCAL VOC 2012, and Camvid, demonstrated that the proposed method can obtain better performance counterparts than other.