Semantic Segmentation Of Road Scene Image Based On Deep Learning

Nowadays,with the rapid development of machine vision technology,the reliability of driverless vehicles has also increased.Machine vision enables the vehicle to perceive the external environment.As an important research direction of machine vision,image semantics segmentation will directly affect the understanding of automobiles to the external scene.Therefore,in order to improve the scene perception ability of driverless vehicles,this thesis does research on the image semantics segmentation algorithm,two semantics segmentation algorithms with higher segmentation accuracy are proposed and implements the application of the two algorithms on embedded platform.The first algorithm is an image semantics segmentation algorithm based on extreme learning machine.Firstly,the image is pre-segmented into many super-pixel blocks by combining the traditional super-pixel segmentation algorithm.Then,the improved region merging algorithm is used to merge the super-pixel blocks to form multiple regions with similar features.Then,four different feature extraction methods are used to extract these super-pixel feature regions.Finally,combining the idea of ensemble learning,we use random combination features to train the extreme learning machine classifier,so as to enhance the generalization ability of the classifier and improve the accuracy of segmentation.Testing on the PASCAL-VOC2012 dataset and the Cityscapes dataset,the experimental results demonstrate that this algorithm can achieve high-precision image semantics segmentation in complex road scenes and compare with the traditional K-means algorithm,the segmentation accuracy is improved greatly.The second algorithm is an image semantics segmentation algorithm based on multi-scale full convolution network,which combines the deep learning idea and uses convolution and pooling to obtain image features.Firstly,the symmetrical network structure of "encoder-decoder" is designed by referring to the SegNet structure.Then,a pooling layer with hollow convolution is added to obtain multi-scale image features.Secondly,a global context structure is designed to solve the problem of loss of image details caused by convolution and pooling operations.Finally,we use the Softmax classifier to classify the extracted image features.Softmax classifier is a multi-class linear classifier,which can give the probability of belonging to each category.Testing on the the Cityscapes dataset,the experimental results demonstrate that the algorithm has higher segmentation accuracy and inference speed than SegNet network.In addition,in order to realize the application of the algorithm on the embedded platform,this thesis simplifies two semantic segmentation algorithms and deploys them on TX2 embedded platform.The experimental results show that the algorithm based on embedded platform is faster than the algorithm based on PC,although it sacrifices a little of the segmentation accuracy.This thesis does research on the problem of image feature extraction in image semantics segmentation task,two more efficient image semantics segmentation algorithms are proposed and further illustrates the practicability of the two algorithms through the transplantation application on embedded platform.