Research On Image Semantic Segmentation Model Of Automatic Driving Based On Residual ESPNet And Fusion Loss Function

In computer vision,a very important branch is image semantic segmentation,which is different from image segmentation and image classification.It is a challenging task in computer vision and one of the key steps to realize scene understanding.With the continuous development of deep learning technology and the continuous improvement of GPU computing power,traditional image semantic segmentation methods are gradually replaced by deep learning.At present,deep learning for image semantic segmentation has become a hot topic for domestic and overseas scholars.In the application scene of image semantic segmentation,automatic driving system is one of the most important application branches.For a complete automatic driving system,it is necessary to accurately and realtime perceive the environment outside the vehicle.External environment perception includes various tasks,especially image semantic segmentation.By means of semantic segmentation of the incoming image from the sensing device,the external contour of each object can be obtained,so that the control system can issue correct and timely control instructions according to various types of structured data analyzed on the driving direction.Due to automatic driving and image semantic segmentation of a high degree of fit,this research mainly around the image semantic segmentation based on deep learning neural network technology.Balance accuracy and speed,so that it can meet the requirements of automatic driving rapidity,and at the same time,each category of objects can get an accurate segmentation contour.This paper presents a model for automatic driving which based on Residual ESPNet.Optimize feature extraction module and add residual structure on ESPNet basic network structure to integrate the underlying semantic information and enrich feature dimensions;By means of changing the convolution form in ESP module,the calculation amount is reduced which will make a great balance between speed and accuracy.At the same time,a loss function based on attention mechanism and level set is proposed.The level set strategy divides the objects into external and internal,and constructs a minimum energy function as the loss of the network,which is used for back propagation to minimize the energy functional and gradually improve the segmentation accuracy.When evaluating the quality of image segmentation,a new evaluation Index based on contour is proposed,which is mainly based on the improvement of BF Score and Jaccard Index.The global statistics of Jaccard Index are used to integrate object contour information to make the evaluation more accurate and objective.