Research On Traffic Scene Segmentation Algorithm Based On Deep Learning

Nowadays,the traffic scene segmentation technology has a wide range of application prospects,especially in the fields of traffic safety,autonomous driving,intelligent transportation,etc.Traditional image processing methods can no longer meet the increasingly complex traffic scene segmentation demands.Deep learning,as a powerful image processing technology,has gradually become the preferred method for exploring traffic scene segmentation algorithms.This paper proposes an improved algorithm based on the Deep Labv3+ model,aiming to improve the accuracy and efficiency of image segmentation in traffic scenes.First,this paper proposes an improved model network based on the Efficient Channel Attention for Deep Convolutional Neural Networks(ECANet)to enhance the segmentation of object boundaries in traffic scenes.The ECANet module is introduced into the Atrous Spatial Pyramid Pooling(ASPP)structure to learn the correlation between different channels and adjust the weights according to the importance of each channel,thus better utilizing features at different scales and avoiding information loss.In addition,the Focal Dice Loss,which combines Focal Loss and Dice Loss with weight,is used to address the issue of sample imbalance.By focusing more on difficult samples and improving the model’s accuracy on difficult samples,the prediction accuracy of the model at the pixel level can be better measured,thereby further improving the model’s classification accuracy and generalization ability.Experimental results show that introducing the ECANet module and using the Focal Dice Loss can improve the segmentation of object boundaries in traffic scenes,and the improved network has an m Io U improvement of 1.15% and an accuracy improvement of 0.4% compared to the original model.Secondly,Attentional Feature Fusion(AFF)is adopted to address the problem of small object recognition loss in semantic segmentation tasks.AFF consists of a Convolutional Block Attention Module(CBAM)composed of spatial and channel attention modules,which can improve the accuracy of small object recognition and ensure the accuracy of object edge segmentation.By learning the correlation between different spatial positions,the weight of each position is calculated to better focus on small objects.Finally,a dual attention module is formed by fusing with the ECANet model to further improve the segmentation performance of the model,and the Focal Dice Loss function is used to solve the problem of sample imbalance.This method provides a more accurate and efficient solution for small object recognition in semantic segmentation tasks.Compared with the original model,the improved network achieved a m Io U improvement of 1.76% and an accuracy improvement of 0.47%.Finally,based on the aforementioned innovations,we combined and fused two novel attention modules to improve the segmentation accuracy of object boundaries and reduce the loss of small objects.To reduce model complexity,we replaced the original Xception backbone with Mobile Netv2,a lightweight network with fewer parameters,faster computation speed,and more suitable for deployment in environments with limited computing resources.Experimental results demonstrate that this method performs well in both segmentation performance and lightweight effect in traffic scenes.On the Cityscapes dataset,our method achieved a 2.54% improvement in m Io U and a0.65% improvement in accuracy compared to the Deep Labv3+ model,while reducing the model size to 43 MB and achieving an FPS of 55.14.