Research On Image Semantic Segmentation Alorithm With Fully Convolutional Networks

Because of the development of deep learning and the emergence of fully convolutional networks,the domain of the image semantic segmentation has been rapidly developed.It is widely used in the fields of driverless,medical diagnosis,machine navigation and so on.Driverless technology has been a research hotspot,in this technology,the perception of the environment around the vehicle is the key points.It can classify images on the pixel-level to obtain the overall information of the image,and the semantic segmentation requires the low-cost vision sensors,so it fits the demand of driverless technology.Fully convolutional networks is a feasible and effective image semantic segmentation algorithm.The algorithm innovatively replaces the fully connected layer with the convolutional layer and applies it to the pixel-level classification task.Deep Lab is an improved algorithm with the fully convolutional networks and this algorithm has a high accuracy.However,there are still some problems in this algorithm,and there is a great space for improvement.We research each sub-module of the algorithm,then research the problem and give the improvement plan to further improve the accuracy of the algorithm.In order to solve the problem that the Deep Lab algorithm does not make full use of global information,resulting in poor results in complex scenes,we introduces the global context information module,this module can provides prior information of complex scenes in the picture,the global context features are extracted and then merged with the local features.This module can improve the expression ability of the features.In order to solve the problem that decoder module of the Deep Lab is too simple and the boundary of the predicted result is rough,we design an efficient decoder module,the shallow layer features are fully utilized,the shallow layer features are merged with the deep layer features,and we adjusts the proportion of the deep features and the shallow features,this way can restore some of the details information,and the boundary of the object is optimized.In order to solve the problem that the Deep Lab is over fitting the fixed size picture,two effective multi-scale feature level fusion modules are designed by combining the idea of integrated learning with the multi-scale model training,and on this basis,an extra supervision module is introduced,this way can improve the robustness of the algorithm.We mainly use the extended Pascal VOC 2012 dataset for experiments.Specifically,first we determine the optimal parameter of the improved method,thenwe fine tune the improved method in the VGG network and the Res Net network.The experimental results show that the three improved method that we proposed in this paper can make the model obtain more expressive features,and improve the accuracy of the algorithm.At the same time,we do some tests on the City Scapes dataset to further verify robustness and effectiveness of the the improved algorithm.Finally,the improved algorithm is applied to the actual scene and has certain practical value.