Multilevel feature fusion dilated convolutional network for semantic segmentation

Recently, convolutional neural network (CNN) has led to significant improvement in the field of computer vision, especially accuracy and speed semantic segmentation tasks, which greatly improved robot scene perception. In this article, we propose a multilevel feature fusion dilated convolution (Refine-DeepLab). By improving space pyramid pooling structure, multiscale hybrid module, captures rich context information effectively alleviates contradiction between receptive size operation. At same time, high-level low-level obtained through multi-level multi-scale extraction can improve capture global performance large-scale target segmentation. The encoder–decoder gradually recovers spatial while capturing information, resulting sharper object boundaries. Extensive experiments verify effectiveness our proposed Refine-DeepLab model, evaluate approaches thoroughly on PASCAL VOC 2012 data set without MS COCO pretraining, achieve state-of-art result 81.73% mean interaction-over-union validate set.