Research Of Semantic Image Segmetation Based On Deep Learning

In recent years,convolutional neural network has been widely used in semantic image segmentation due to its good performance.However,semantic segmentation networks with high performance usually have complex network structure with a large number of network parameters,witch required a lengthy computation,and it is difficult to be applied to applications with high real-time requirements such as embedded devices and mobile devices.In this paper,the number of parameters,running speed and performance of the network are taken into consideration comprehensively.We proposed a lightweight multi-level features cascade semantic image segmentation network.Specifically,the main contributions of this method are as follows:In this paper,a lightweight multi-level features cascade semantic image segmentation network is proposed.Usually,the high performance networks' structure have wide width and deep depth,which have large number of parameters and large amount of computation.We fine tuning the lightweight classified network as the feature extraction structure of our network,which improves the shortcomings caused by big computation numbers.In addition,in order to reduce the size of the network scale and ensure the segmentation accuracy of the network,the atrous residual feature refine module(AR)and the deep atrous spatial pyramid pooling module(DASPP)were also proposed.AR module is composed of two residual branches,which are used to enhance the edge contour information of objects of different sizes in shallow features.The DASPP module is composed of atrous convolution with different receptive field sizes,which is used to enhance the semantic information of different size target objects in depth features.Compared with the existing semantic segmentation algorithms,the lightweight multi-level features cascade semantic image segmentation network proposed in this paper,has better semantic segmentation accuracy and real-time performance,and can be better applied to mobile devices and embedded devices with real-time requirements while ensuring segmentation accuracy.