Research On Semantic Segmentation Of Road Scene Images Based On Deep Learning

Recently,with the development of deep learning,the image semantic segmentation based on convolutional neural network has been widely used in various fields,among which autonomous driving is one of them.Semantic segmentation of road scene image,as the most important component of autonomous driving,has important practical significance.However,road images are different from ordinary images in that they contain complex and different scales of targets,so semantic segmentation still faces challenges.The main problems are the slow segmentation speed and insufficient feature selection ability.To address this problem,this paper studies the semantic segmentation of road scene images based on the relevant knowledge of deep learning,and two improved network models are proposed.The specific work and innovations are as follows.(1)A multi-scale road scene semantic segmentation network based on attention mechanism is proposed,which is a new real-time semantic segmentation network improved from Link Net.Firstly,a new position-channel attention mechanism is proposed,which focuses on the saliency information from the two dimensions of the position and the channel of the road image,so as to increase the ability of the network to extract image features.Secondly,an atrous spatial pyramid pooling is added between the encoder and the decoder,which uses parallel atrous convolution and pooling to extract multi-scale features,capturing and outputting more refined results.The performance of the proposed model has been improved by validating on the Cityscapes dataset.Compared with the original model,MIo U is improved by 5.01%,and MPA is 93.42%.Under the same experimental conditions,the number of parameters has reduced,the operation speed is significantly improved,and the segmentation accuracy is optimal,compared with the classical segmentation model.(2)A saliency-assisted collaborative learning network for road scene semantic segmentation is proposed,which is an improved collaborative semantic segmentation network based on deeplabv3+.Firstly,a depthwise separable convolutional spatial pyramid pooling model is proposed,which consists of multiple depthwise separable convolutions and pooling operations in parallel to form a pyramid shape.Among them,depthwise separable convolution can effectively reduce the number of parameters,and parallel combination can effectively capture contextual information.Secondly,the idea of collaborative learning is adopted,combined with the saliency detection task,and information complementation is achieved through convolutional layer sharing,which provides effective information for segmentation.Finally,the losses of the two tasks are processed with homoscedastic uncertainty,and different weight information is given to the losses of the two tasks to optimize the loss function of the cooperative task.After training and testing experiments,the proposed network achieves MIo U of 78.94% on the Cityscapes dataset,and MIo U of 60.90% on the PAC VOC 2012 dataset,which are both improved compared with the current advanced algorithms.The segmentation results show that both models proposed in this paper effectively improve the segmentation accuracy,and the models are more complete and clearer for the recognition of fine objects and boundaries,and obtain excellent results in the semantic segmentation of road images.