Research On Image Semantic Segmentation Based On Dense Deconvolution Aggregating Network

Semantic segmentation plays an important role in robot vision and can be utilized in many real-world applications,such as virtual/augmented reality,robotics,and self-driving car.Recently,exploring multiple feature maps from different layers in fully convolutional networks(FCNs)has gained substantial attention to capture context information for semantic segmentation.This paper proposes two tasks for semantic segmentation:(1)In this paper,a novel encoder-decoder architecture has been proposed,called dense deconvolution aggregating network(DDANet),for semantic segmentation,where the contextual cues are investigated via densely usampling the convolutional features of deep layer to the shallow deconvolutional layers.The fused feature maps complement each other,enabling the network to more fully explore the global information contained in the image.The proposed DDANet is trained in terms of end-to-end segmentation to match the resolution of input image.This paper evaluate the DDANet on self-driving dataset City Scapes.The experimental results show that DDANet outperforms recent FCNs and encoder-decoder networks(EDNs).DDANet also achieve superior results on other two benchmarks: PASCAL VOC 2012 and ISBI 2012 dataset,for the challenges of indoor/outdoor scene understanding and biological segmentation.(2)This paper continues to make improvements to DDANet,and proposes R-DDANet(Refine Dense Deconvolution Aggregating Network).This network as a whole is also an encoder-decoder structure.In the encoding side,R-DDANet replaced the previous VGG-16 with ResNet-34.The MSC module is newly added to the decoder.The feature map at the same layer of the encoder is copied from the decoder directly to the decoder instead of being sent to the decoder through the MSC module,so that it can learn context information of different scales at the encoder.This paper verified on two datasets,Cityscapes and PASCAL VOC2012.Experiments prove that the improvement in R-DDANet is indeed effective.