Research On Semantic Segmentation Based On DeepLab With Propagating Deep Aggregation And Boundary Refinement

Semantic segmentation is a very significant part of Computer Vision field,and plays an important role in the subsequence work of image processing.Semantic segmentation is a process of segmenting digital images into multiple modules,which is composed of pixels or super-pixels,and giving each pixel real semantic label.The purpose of semantic segmentation is to simplify image and to endow more significance with image,making it easier to analysis.At present,the Fully Convolution Neural Network is the mainstream of semantic segmentation tasks,and many excellent algorithms have been derived.Deep Lab model is a relatively mature architecture among those algorithms.However,Deep Lab does not take into account that the information contained in each layer of the network is meaningful,only for specific layers of feature fusion,and the loss of information is serious in the process of propagation in the shallow layer,as well as the use of fully connected CRF occupies too much computing resources and too long computation time.According to above research,this paper aims at improving the accuracy of image semantic segmentation and propose a simple and effective algorithm for obtaining greater global context and a boundary refinement with simple structure and without adding learning parameters which is based on the version 2 of Deep Lab model.The details are as follows:(1)Semantic segmentation method based on Deep Lab model with Propagating Deep Aggregation structure overcomes the problem that the network can't aggregate all feature channels and layers and lose the information seriously in the process of propagation of feature information from shallow layers.The paper designs an aggregation structure which can connect all convolutional blocks and greatly guarantee the integrity of shallow feature information in the propagating process.The structure is consisted of two parts,one is hierarchical deep aggregation which combine deeper and shallower convolution blocks to learn more feature combinations across layers;the other is iterative deep aggregation,which divides the convolutional network into several stages depended on the output resolution of convolutional layers,starting from shallow stage to deep aggregation step by step,and finally updating the output node iteratively.In the stage,using hierarchical deep aggregation to fusing blocks.Between the stages using iterative deep aggregation to fusing all stages and update output.In this way,all convolutional block features can be fused and shallow information integrity can be guaranteed in the dissemination process.Finally,the classical data sets like PASCAL VOC 2012 and PASCAL-Context are used to test,and the feasibility and validity of the proposed algorithm are verified by comparing with other excellent semantic segmentation algorithms.(2)The semantic segmentation method based on Deep Lab with boundary refinement block,which is different with other mainstream methods of boundary treatment such as using contour detection on convolution neural network or probabilistic graph model to connect convolutional neural networks.In this paper,a boundary refinement module based on residual network is designed to refine image boundary.The boundary refinement module consists of two branches,one of which transmits rough feature mapping,and the other uses the residual between the expected value and the actual output to find the boundary position and get the segmentation map which focuses on learning fine edge information.The output position of each convolutional layer in the network is connected with a boundary refinement module.Referring to the idea of jump connection,the results of each layer of boundary refinement module are fused from deep to shallow,and the precise location of the pixels in the image boundary area is enhanced by using the boundary information contained in each convolution layer.In the end,the feasibility and effectiveness of the algorithm we proposed are verified by designing a set of comparative experiments and using multiple sets of dataset.