Real-time Semantic Segmentation Of Road Driving Scenes Based On Attention Mechanism

Semantic segmentation,a popular research branch in computer vision,has important applications in areas such as unmanned vehicles and medical image segmentation.Semantic segmentation is a pixel-level classification task,which classifies the input image pixel by pixel.In the unmanned driving task,road scene detection is what needs to be focused on.Real-time semantic segmentation can segment road driving scenes to help driverless systems understand road condition information and make timely decisions.Traditional autonomous driving technology first collects data from the driving environment through sensors,then uses algorithms to analyze the data,and finally controls the vehicle according to the analysis results.In contrast,the semantic segmentation algorithm based on deep learning takes the collected images and passes them through a convolutional neural network to extract the features contained in the images and perform end-to-end image segmentation.Compared with traditional algorithms,the deep learning-based semantic segmentation algorithm has the advantages of high segmentation accuracy,end-to-end execution,and higher real-time performance.As semantic segmentation networks are widely used for practical tasks,various problems have gradually surfaced.For example,for objects with complex features,the network may not be able to learn their features sufficiently,leading to unsatisfactory segmentation results of the overall network.In addition,the lack of sensitivity to the information of multi-scale objects may also lead to poor segmentation results.To solve the above problems,this paper starts from the road driving scenario and firstly applies the attention mechanism to learn objects containing complex features,and then uses large perceptual fields to improve the network’s ability to recognize fine objects,thus improving the segmentation accuracy of the network.The main research contents of this paper are as follows.（1）In this paper,we propose a Real-Time Semantic Segmentation Network Based on Feature Coordination Attention（FCANet）,aiming to improve the feature learning capability of the network for complex objects in images.In this paper,the network adopts a dual-path structure consisting of spatial and semantic paths,where the spatial path is used to extract rich spatial information and the semantic path is used to extract deep semantic information.In addition,this paper introduces the attention module and designs a multi-path attention mechanism and a feature coordination layer.Multi-Channel Attention Mechanism（MCAM）and Feature Coordination Layer（FCL）are designed to establish the relationship between channels of the semantic information feature map and then refine the relationship between channels.The Feature Coordination Layer fuses the bottom semantic information in the semantic path with the top semantic information and spatial information respectively.Through experiments,the network proposed in this paper achieves 62.4% MIo U in the test set of Cam Vid and is capable of segmentation speed of 116 fps,which demonstrates the effectiveness of the multiplex attention mechanism and the superiority of the feature coordination layer.（2）In this paper,we propose Real-time Semantic Segmentation Network with Multi-Scale Multi-Channel Attention（MSMANet）,which aims to improve the segmentation capability of the network for fine objects in images.In this paper,we further optimize the segmentation capability of feature coordinated attention network by proposing Extended Acceptance Domain Module（EADM）.A large receptive field enables the network to extract intricate relational features between objects and facilitates the network to recognize fine objects.The network is trained and tested on the Cityscapes test set,and the segmentation accuracy on the test set is 81.0%,which demonstrates that the proposed Extended Acceptance Domain Module can improve the segmentation effect of the network on fine objects.