Research On Real-time Semantic Segmentation Technology Of Traffic Scene Images

With the development of deep learning technology,image semantic segmentation has shown practical application value in areas such as autonomous driving,augmented reality,and video surveillance.For example,in the field of autonomous driving,semantic segmentation algorithms segment and locate the scene where the vehicle is located,thereby improving the safety of autonomous driving.However,the existing mainstream methods usually build deeper and wider convolutional neural networks in order to obtain the ultimate segmentation accuracy.It is difficult to apply to the actual scenes with limited resources by deepening and widening the network.In actual application scenarios,there is an urgent need for accurate and efficient segmentation technology.This paper uses a deep learning-based method to study the semantic segmentation task of traffic scenes from three aspects,as follows:Aiming at the low efficiency of the existing attention mechanism network on the task of semantic segmentation of traffic scenes,a lightweight semantic segmentation algorithm based on dual attention is designed and implemented.Through deep separable convolution,a lightweight dual attention structure is constructed and the parameters are simplified,which improves the efficiency of the algorithm.At the same time,a multi-scale attention pyramid module is designed to use feature maps of different receptive fields or output prediction results at different scales to improve the robustness to complex backgrounds in traffic scenes.Lastly,an adaptive multi-scale prediction fusion module is designed to adaptively fuse the prediction results of multiple different receptive fields,thereby further enhancing the network prediction ability.The results of ablation experiments on the Cityscapes dataset show that the proposed method has obvious advantages in segmentation performance,segmentation efficiency and the number of parameters.In order to reduce the dependence of the segmentation model on the pixel-level annotation data of traffic scenes,a semi-supervised semantic segmentation algorithm model based on knowledge distillation is designed and implemented.Through the selfcorrecting module,weakly labeled data is iteratively optimized and pseudo-labels are generated;a multi-student collaborative learning method is designed to improve the ability of the student network to accept potential knowledge;and the knowledge distillation structure migration semantics of the teacher-student network is introduced Structured information.The results of ablation experiments on the Cityscapes and Cam Vid datasets show that this method can improve the predictive performance of the student network,and at the same time generate refined pseudo-labels for the Cityscapes weakly labeled data,and solve the problem of insufficient refined label samples in the Cityscapes dataset.The network performance obtained by training with the original label data combined with the pseudo label data can be further improved.In response to the real-time requirements of semantic segmentation tasks in traffic scenes,a real-time semantic segmentation algorithm based on mixed-depth separable convolutions is designed and implemented.It integrates local features through mixeddepth separable convolutions and enhances the expression ability of multi-scale target features.At the same time,feature rearrangement is used to eliminate the problem that information cannot be exchanged between channels.And designed two different nonbottleneck structures and residual up-sampling modules.A variety of ablation experiments were carried out on the PASCAL VOC2012 and Cityscapes data sets and the effectiveness of each module was verified.Finally,the results of training and testing on the Cityscapes dataset show that the proposed method achieves a good balance between segmentation efficiency and segmentation accuracy,and the algorithm performance is better than the current mainstream real-time semantic segmentation algorithms.