Semantic Segmentation Of Mountain Road Image Based On Deep Learning

Image segmentation is widely used in the field of computer vision.With the emergence of deep learning,image semantic segmentation has also been widely studied.Semantic segmentation technology is a combination of target classification and image segmentation,its role is to divide the image into different pixel level areas.Semantic segmentation is widely used in pedestrian detection,medical influence segmentation,automatic driving and other fields.At present,most of the research is about urban road scene,but this paper focuses on the semantic segmentation in mountain road scene.In this paper,the semantic segmentation algorithm and the commonly used network structure are described in detail,and three different network structures are used for the experimental comparison of the mountain road scene.Corresponding to the data collected in the mountain road scene,the semantic segmentation data set of the mountain road scene image is made,which contains 8 categories,which are mountains,cars,pedestrians,cyclists,roads,signs,caves and background.According to several evaluation indexes,the performance of different semantic segmentation model networks is compared.Firstly,the UNet based on encoding and decoding structure is designed and constructed.Its advantage is that it does not need a large number of marked data samples,a small number of samples can achieve a better effect,and the structure is simple and easy to design.When the UNet was tested in the mountain road data set,the m Io U was 61.31% and the detection speed was 8.78 fps.Real-time lightweight semantic segmentation is also becoming increasingly important in autonomous driving scenarios.Most of the semantic segmentation network models have a slow prediction speed.In order to solve this problem,a fast semantic segmentation model ICNet based on image cascading network is used in this paper.The model combines the pyramid pooling structure and the image cascading frame,fuses the features of low resolution fast detection speed with the features of high resolution feature details,and improves the detection speed.The ICNet tested m Io U in the mountain road data set at 53.08%,the speed reached 21.15 fps,close to the real-time processing requirements of the algorithm.The detection speed of ICNet is 2.41 times higher than that of UNet.In order to improve the segmentation accuracy of network in mountain road test,the UNet structure was improved.Since the UNet structure has no feature selection ability,an attention mechanism is introduced to process specific areas.Meanwhile,in order to deepen the structure of the network,the trunk network of the UNet coding structure is replaced by a residual structure.In addition,the attention mechanism cannot be superimposed randomly in the network structure,and the attention mechanism will continuously decrease the value of the feature graph.To solve this problem,this paper proposes an attention-guided residue-semantic segmentation network(RAUNet)to improve the segmentation accuracy of test sets.Compared with UNet and ICNet,the RAUNet model improved the m Io U on the mountain road data set,reaching 69.79%,and the detection speed was 5.14 fps.