Research On Semantic Segmentation Algorithm Of Traffic Scene Image Based On Deep Learning

With the extensive and profound application of deep learning in the field of artificial intelligence,it plays a crucial role in the control fields such as machine vision and robotics.By using computers instead of humans to extract visual features from images and videos,perform intelligent training and network testing.The main applications are image recognition and classification,target detection and tracking,and image semantic segmentation,etc.The main research content of this paper is semantic segmentation algorithm of traffic scene image based on deep learning.Image semantic segmentation can be applied to automatic driving control,intelligent assisted driving,intelligent robot and other control fields.In automatic driving,the main task is to recognize and segment the pixels of various objects in traffic scene images.In the process of segmentation,semantic feature information of objects such as roads and vehicles is extracted.To further understand the overall content of the image to help intelligent driving vehicles choose the optimal driving route.As follows:(1)Aiming at the low segmentation accuracy and incomplete segmentation edge of the full convolutional network for small targets,a vehicle scene semantic segmentation algorithm based on separable convolutional residual network was proposed.Firstly,a more comprehensive edge feature extraction is carried out by using separable residual network blocks.Then the feature graph is restored by deconvolution to obtain the prediction segmentation result graph.In the Camvid test set,Miou increased from 76.85% to 83.30% compared with the full convolutional network.(2)Aiming at the depth of feature extraction network of full convolutional network and the resolution of feature image after pooling,a semantic image segmentation network based on feature extraction of bilevel residual network DResnet was proposed.Firstly,a two-layer residual network is built to extract different features of images,so that the fusion features have accurate perception ability.Secondly,the fusion starts from feature graph X1,and uses 2x deconvolution method to fuse features of different levels.The branch training method is used to train the network and improve the training accuracy of the network.In Camvid,compared with full convolutional network,Miou increased from 49.72% to 59.44%,and in Cityscapes,it increased from 44.35% to47.77%.(3)In view of the weak perception ability of some existing semantic segmentation models to different scale targets of training images,the feature image resolution after pooling operation is small or even feature loss,an image semantic segmentation algorithm Ms Res Net based on multi-scale residual and jump feature splicing is proposed.Firstly,a multi-scale feature extraction module is constructed to extract the features from the input image.Secondly,the feature images output by five layers are up-sampled by jumping feature fusion and tensor splicing.Finally,a multi-branch training method is proposed,which can be selective for training different branches of the network.In Camvid,compared with the full convolutional network,Miou increased from 49.7% to63.0%,and 3.56% higher than DResnet network.In Cityscapes,Miou increased from 44.3% to 50.7%,2.93% higher than DResnet.