Semantic Segmentation Of Road Scene Images Based On Convolutional Neural Networks

Computer vision is one of the important research directions and breakthroughs in the field of artificial intelligence.Image semantics segmentation technology is a hot research direction in the field of computer vision.Image segmentation technology is to classify the input image pixels by pixels,expand the image classification to the classification of pixels,and then apply it to various fields of computer vision.Image segmentation technology has been widely used in image retrieval,video tracking,target recognition and robot navigation and other practical visual application scenarios.Image segmentation technology is very important to solve the problem of robot vision.Current image semantics segmentation algorithms can generally be divided into two categories: traditional segmentation methods and deep learning-based semantics segmentation methods.Aiming at the requirement of unmanned driving for semantics segmentation task of road scene image,this paper studies semantics segmentation based on depth learning method,especially for accurately segmenting the more important target of unmanned driving safety in complex road scene.In image semantics segmentation using depth learning method,convolution neural network structure is usually used to extract multi-level feature map,and then the method of up-sampling is used to restore the feature map to the original image size of the segmentation results,so as to achieve the purpose of pixel-by-pixel classification.The main work of this paper includes:(1)By dividing all kinds of targets in road scene into importance levels,an image segmentation algorithm based on importance weighting is designed,which can make the model more accurately segment the targets that are critical to the safety of unmanned driving in road scene,and at the same time,it has limited influence on the segmentation accuracy of the target area with low importance level.This paper uses City Scapes data set and Camvid data set to train the network model optimized by the proposed algorithm,and verifies the validity and generalization ability of the method.(2)The scoring mechanism of neighborhood category correlation and the network structure of cross-layer feature fusion are proposed.Meanwhile,migration learning and fine-tuning are used to optimize the image semantics segmentation of road scenes.Generally,classical convolution neural network modules such as VGG16 and their pre-training parameters on Image Net are used as the front-end of the model to extract image features.In order to apply these classical convolution modules to the task of road scene semantics segmentation,an improved version of Res Net101,namely Res Net101(V2),is adopted,which combines with neighborhood category correlation scoring mechanism and cross-layer feature fusion mechanism.Several convolution layers in front of the convolution part of the neural network architecture are frozen,and other convolution layers are thawed,and then trained.Compared with the training of random initialization parameters,the training time can be greatly reduced,and the final output results will be relatively good.The purpose of migration learning and finetuning is to optimize the coder(convolution layer)of the neural network structure used in image semantics segmentation by neighborhood category correlation scoring mechanism,and verify their performance in image semantics segmentation of road scenes.The evaluation of this performance includes segmentation accuracy and Mianyou index.(3)Conditional random field has a good effect in optimizing the smoothness of image segmentation and boundary processing.The scoring mechanism of neighborhood category correlation and cross-layer feature fusion structure proposed in this paper also have a good effect on boundary processing after some parameter adjustment.In order to verify the effectiveness of boundary treatment of the structure proposed in this paper.On the basis of importance weighting algorithm,three groups of comparative experiments are conducted to analyze the experimental results.1)Segmentation of image semantics using only neighborhood category correlation scoring mechanism and cross-layer feature fusion structure.2)Conditional random fields are used only for image semantics segmentation.3)At the same time,two methods are used for image semantics segmentation.Compare the results of the three groups of experiments in order to select the appropriate method under the appropriate situation.