Research On Road Scene Segmentation Algorithm Based On Convolutional Neural Network

Instance segmentation is a computer vision task that is closest to human real visual experience.It is a deep understanding of image scene and has high application value in the field of autonomous driving.However,the actual road scene is complex,changeable and messy,and the segmentation target is usually multi-level,and will show different shapes due to changes in light and shooting position,which will increase the difficulty of actual road scene segmentation.Therefore,the road scene segmentation algorithm based on convolutional neural network studied in this paper has important practical significance for the construction of smart transportation.The main research contents are as follows:(1)Selection and preprocessing of data set.For road scene segmentation models,the diversity and richness of training data directly determine the performance of the model.By comparing the geographic area,data label density,the number of labeled pixels of each category,and scene complexity of several road scene data sets,this paper selects the Cityscapes data set that is closest to the real road scene,and performs data preprocessing operations on the dataset,such as data cleaning,image sharpening,brightness processing,and data enhancement,to enrich the number and diversity of images in the dataset and provide data guarantee for subsequent road scene instance segmentation algorithm research.(2)Research on road scene segmentation algorithm based on Mask RCNN.In view of the fact that when the Mask RCNN algorithm is directly used for road scene segmentation,the accuracy of small target segmentation is not high and overlapping targets are missed.This paper uses a new multi-scale backbone structure Res2 Net to replace the Res Net network,which increases the receptive field range of the model and improves the multi-scale feature extraction capability of the model without increasing the computational load.The Soft-NMS algorithm is used to replace the NMS algorithm,which effectively reduces the missed detection rate without increasing the complexity of the algorithm.A combination of step-by-step training and end-to-end training is adopted to further improve the segmentation effect of the model.The experimental results show that the optimization method proposed in this paper can solve the defects of the Mask RCNN algorithm in road scene segmentation to a certain extent,and has a certain degree of competitiveness in the current advanced algorithms.(3)Research on road scene segmentation algorithm based on improved Center Mask.When the Center Mask algorithm is directly used for road scene segmentation,there will be problems of low segmentation accuracy and difficulty in segmentation of small targets.This paper designs a new network branch SCAG-Mask,which increases the attention mechanism in both the channel and spatial dimensions.Pay attention to meaningful pixels in space and suppress meaningless pixels,while re-weighting the channels to enhance useful feature channels and suppress useless feature channels.This paper also proposes an improved regression strategy.On the basis of the original regression strategy,the prediction of grid vertices is added to form a prediction method that combines the center point and vertices of the grid.Then according to the modification of the regression strategy,the new loss caused by the corner point is also added to the loss function.The experimental results show that the optimization method of the Center Mask algorithm in this paper will sacrifice less speed in exchange for better segmentation accuracy.Compared with the optimized Mask RCNN experimental results,the average segmentation accuracy of the improved Mask RCNN is higher than that of the improved Center Mask,but the optimized Center Mask segmentation speed is faster,and the segmentation accuracy in large targets is slightly better than the improved Mask RCNN.