Research On Real-time Semantic Segmentation Algorithm Based On Deep Learning

Image semantic segmentation is to classify each pixel in the picture.It is a pixel-level classification task and a research hotspot in computer vision.It has great application value in the fields of autonomous driving,video surveillance,and medical treatment.With the development of convolutional neural networks,the field of semantic segmentation based on deep learning has also achieved great development,resulting in many high-precision semantic segmentation algorithms.Although many high-precision semantic segmentation algorithms have been produced,these high-precision algorithms have a huge architecture and a large number of parameters,making these high-precision algorithms very slow and unable to be applied to practical applications.However,the existing fast real-time semantic segmentation algorithms all have the problems of insufficient receptive field and insufficient spatial details,which leads to their low accuracy and cannot guarantee the effect of actual application.In order to solve this problem,this paper proposes a dilated pyramid attention module to increase the receptive field of the network and capture the multi-scale information of the object,and use the attention mechanism to improve the accuracy of the network.This paper uses the dilated pyramid attention module to construct a real-time semantic segmentation network: the dilated pyramid attention network,In order to further improve the accuracy of the network,this paper proposes an refinement module that uses the original image to refine the results generated by the dilated pyramid attention network,which further improves the accuracy of the network.The main work of this paper is as follows:1)To propose a real-time semantic segmentation algorithm based on the dilated pyramid attention.Aiming at the problem of small receptive field and low accuracy in real-time semantic segmentation algorithms,this paper uses the dilated pyramid structure to capture the multi-scale information of objects,increases the receptive field of the network,and introduces the attention mechanism,and proposes a dilated pyramid attention module.To use the dilated pyramid attention module and the backbone network Res Net18 construct a real-time semantic segmentation network:the dilated pyramid attention network.Experiments on the Cityscapes dataset show that the high efficiency of the dilated pyramid attention module proposed in this paper and the superiority of the dilated pyramid attention network.2)To propose a real-time semantic segmentation algorithm based on the result refinement.In order to further improve the accuracy of the dilated pyramid attention network,this paper refines the results produced by the dilated pyramid attention network,this paper proposes an refinement module that uses original images of different scales to refine the results,supplements spatial details,and improves accuracy.This paper combines the refinement module and the dilated pyramid attention network to construct a real-time semantic segmentation network: the result refine network.Experimental results show that the proposed refinement module has significant effects and greatly improves the accuracy of the network.