Research On Semantic Segmentation Of Cityscape Based On Unsupervised Domain Adaptation

Semantic segmentation is an significant research in the field of computer vision,it is the basis of scene understanding and analysis,and widely applied on automatic driving,human-computer interaction,computational photography,image search engine,augmented reality,etc.Recent years,with the development of deep learning,recent deep networks achieved state of the art performance in semantic segmentation task.Unsupervised domain adaptation exploit unlabeled target data,hence the application in semantic segmentation becomes wider,but the existing unsupervised domain adaptive approach for the domain gap between source and target data are still not well aligned the distribution between the two domains,the edge of the object segmentation is coarse,and segmentation accuracy and generalization performance remains to be improved.As to above problems,the mainly contribution of this paper are following two aspects:(1)AdaptSegNet(Domain Adaptive Segmentation Network)achieves better results on semantic segmentation,but this model use GAT5 and CityScapes directly for adversarial training,which easily leads to the adversarial loss during the training producing gradient explosion in back propagation,then causes model’s performance degradation.To address this issue,domain adaptive adversarial segmentation network which combining with FPS(Fast Photo Style)style translation model is proposed,and also applied to the segmentation of city scene.Firstly,the algorithm adopt FPS model to style translation on GTA5,then obtained the new dataset FPS＿GTA5,which get more closer on color and texture comparing with CityScapes dataset,and use FPS＿GTA5 dataset replaces source dataset GAT5 in AdaptSegNet;Secondly,so as to verify the performance of FPS model,using Cycle Gan model takes the place of FPS model to conduct same operation to generate Cycle＿GTA5 dataset,meanwhile,making comparison and analysis on the visual effect and similarity measure of FPS＿GTA5 dataset,Cycle＿GTA5 dataset,and CityScapes dataset;Finally,in order to certify the method referred by this paper,we use GAT5 and SYNTHIA as source datasets,CityScapes as target datasets to conduct experiments,furthermore,make comparison with current several mainstream unsupervised segmentation algorithm.The experiment result shows that: from GAT5 to CityScapes experiment,the model of this paper compares with AdaptSegNet,the segmentation precision has increased above 7% on Wall,Pole,Sign,Terrain,Car and Bus,improved 2.7% on mIoU(mean Intersection over Union).As to the experiment from SYNTHIA to CityScape,the model of this paper compares with AdaptSegNet,segmentation accuracy has increased above 3% on Road,Sidewalk,Car and Bus,improved 1% on mIoU.(2)AdaptSegNet adopts the segmentation map between source and target to train with adversarial learning,If there is a large domain gap between the source and target,it will cause a gradient explosion due to the large adversarial loss,and the model’s segmentation performance will deteriorate,so the segmentation performance of the model still need to be promoted.To solve this problem,proposing to add Domain adaptive adversarial segmentation network of self-training module based on AdaptSegNet.The algorithm introduce the self-training module into adversarial learning,this can improve the capability of segmentation model;Since the self-training module adopt Cross entropy loss,which easily ignore the segmentation with low probability category,thus result in the issue of unbalanced probability.This paper use Maximum square loss to replace Cross entropy loss,which can alleviate the issue effectively that low probability category was ignored then improve accuracy;Moreover,aiming at balancing the category of high probability and low probability further,we introduce the class balance weighting factor into the maximum square loss,thus can better alleviate the problem of unbalanced probability and further improve accuracy;Finally,we evaluate on the CityScapes validation set with 500 pictures,and conduct the ablation experiment,also compare with several mainstream methods.The experiment result shows,comparing with AdaptSegNet model,the method of this paper segmentation accuracy has increased 3.1%,7.8%,8.8%,2.8% on Road,Sign,Car,Bus respectively,also 1.1% improved on mIoU.