Research On Image Semantic Segmentation Algorithm Of Traffic Intersection Based On Fully Convolutional Neural Network

Currently,most deep convolutional neural networks used for image semantic segmentation face many issues,particularly relevant in the fields of automatic driving and medical imaging: the number of front-end feature extraction network parameters is large,the calculation is huge,and the efficiency of the algorithm cannot be guaranteed.Additionally,the generalization ability of the segmentation model is insufficient,and the output segmentation map is incomplete and unclear in its contour details.Clearly,designing an efficient,high-performance lightweight neural network is the key to solving the problem.Based on further research of more mature image semantic segmentation algorithms,the fully convolutional neural network Deep Lab V3 Plus algorithm was selected as the basic neural network for the experiment,which attempts to make the original algorithm more efficient,enhance the ability of model generalization,improve segmentation,and optimize and improve the algorithm for image sharpness.The experimental results were then compared with actual scene detection.The backbone network of the original algorithm Deep Lab V3 Plus was improved in the following ways in order to better the timeliness and accuracy of the original algorithm.The model using Xception is better in operational accuracy,and the model using Resnet has a shorter average output time of the segmentation graph.The model for Xception was changed to a lightweight neural network Mobile Net V2 and the decoder module was removed both to increase speed and to reduce the number of parameters.The model of the backbone network was changed from Resnet to improved Resnet,and more non-linear operations were added to improve the learning ability of the model.In terms of enhancing the generalization ability of the model,the data enhancement operation was performed on the experimental data set,so that the generalization ability of the model is enhanced,and the possibility of overfitting the model is reduced.In improving the timeliness of the algorithm and improving the sharpness of the segmentation map,the fully connected conditional random field is used in post-processing to perform multiple iterative optimizations on the segmentation map to obtain a more complete and clearer segmentation map in the contour details.Ablation experiments and parameter analysis show that the improvement and optimization of image semantic segmentation algorithms based on fully convolutional neural networks can achieve a clearer segmentation map while improving the timeliness of the algorithm and enhancing the generalization ability of the model.;The mean intersection of union of accuracy evaluation indicators is better than most previous algorithms.Finally,the backbone network is derived from a lightweight neural network Mobile Net V2 model with the decoder module removed,in addition to a fully connected conditional random field as a post-processing package into a complete model,which is applied to traffic intersections.In the actual scene,it has certain social practical value.