Research On The Street Scene Semantic Segmentation Technology Robust For Multiple Scenes

With the development of deep learning technology,the topics such as smart city,intelligent machine and self-driving are rising.Street semantic segmentation result as the start point for these applications has become a research hotspot.Existing methods have a good performance under the sunny day and normal lighting conditions,but it can't perform well under rainy and night.It's therefore a key technology how to obtain a robust semantic segmentation model for multiple scenes.The aim of this paper is to study a semantic segmentation model robust for multiple scenes,such as sunny,rainy and night.It's able to alleviate the existing problems in the street scene semantic segmentation to a certain extent.The main contributions are as follow:Design a lightweight semantic segmentation model: In this paper,I design two building blocks called multi-scale feature extracting structure and cascaded gradient propagation structure respectively,which can extract the more complete feature from the image and end up giving semantic segmentation model a better performance.Design a module that can generate the night scene and rainy scene image at the same time: In this paper,I introduce a data augmentation module that can generate rainy scene image and night scene image at the same time,which can be inserted into the network during the training phrase.It solves the problem of performance degradation in noise scenes such as rainy and night.Design a Group MMD loss: In this paper,I design a Group MMD loss to eliminate the distribution different between scenes,which make model learn the more common feature from image.