Semantic Segmentation In Complexed Environment Using Deep Nerual Network

Semantic segmentation is one of the most challenging and important research issues in the field of intelligent driving.Using depth learning model to deal with semantic segmentation task has become the mainstream in the academic and application fields.For semantic segmentation task,the performance of deep learning model can be significantly improved by investigating the two-dimensional sequence relationship of images in the process of image modeling.On this basis,non local module and other forms of attention module are proposed.The main contents of this thesis are as follows:1.Aiming at the problem of image sequence relationship modeling,the attention mechanism is introduced,and the global relevance is constructed by using non local modules.Meanwhile,the lightweight skeleton network of mobilenetv3 is used to propose a semantic segmentation model with fewer parameters but similar modeling ability.Experimental results show that the proposed model improves the pixel level accuracy of 9.24 and 0.10compared with the baseline model.2.Aiming at the problem of convergence rate caused by the traditional pixel classification loss,this thesis uses the combination of global classification loss and pixel classification loss to speed up the convergence rate of the model.Considering that the generalization ability of the model can be improved to a certain extent by adding margin into the classification loss,this thesis adds the amsoftnax loss function to the overall loss function.The simulation results show that the proposed method produces a certain performance gain for the proposed model,and achieves the accuracy of 0.89%and 72%respectively in the average PAPR and pixel accuracy index.