Research On Road Detection Based On Attribute Converter And Weak Supervision

Road detection is the focus of the field of artificial intelligence,and road detection at night is difficult because of its highly similar samples and fuzzy road edges.Semantic segmentation based on depth learning is the main method of road detection.Semantic segmentation is to achieve the goal of road detection by classifying each pixel of the road scene.The mainstream semantic segmentation networks include PSPNet,SENet and SETR.PSPNet proposes a pyramid pooling module,which obtains context information through pooling layers of different scales;SENet proposes a channel attention method based on global pooling compression,which improves the network’s ability to extract global information;Based on VIT(Vision Transformer),SETR cuts the picture into several small pieces and sends them to the pure attention codec for training.The above methods have made great achievements in the field of road detection,but there are still some shortcomings;For example,PSPNet has insufficient research on global information extraction;The channel attention method of SENet is based on the global pooling compression,which leads to excessive loss of semantic information;The SETR model based on pure attention is too large and the training time is too long,which is not suitable for small data sets.In view of the insufficient number and diversity of training sets,the current mainstream method is to enhance data sets.However,data enhancement can be divided into traditional methods and generation antagonism(GAN).Data enhancement based on traditional methods mainly involves geometric transformation of images and increasing contrast,which can improve the generalization performance of the network to a certain extent,but does not solve the fundamental problem of insufficient diversity.GAN is one of the most important and worth studying image enhancement techniques,but it has the problems of long training time,image similarity and low feature utilization.Moreover,the data set expanded by GAN has no calibration information,which also restricts the development of generation countermeasure network applied in data enhancement field.In view of the above problems,the research work of this paper is as follows:1.Combine the generation network with the detection network,use discriminator to optimize the features,and add interference features in the training process to improve the detection accuracy.In view of the problem that the lack of edge constraint information in the attribute conversion process leads to the geometric deformation of the road edge and the low quality of the generated image,the daytime road image and its edge information are sent to the attribute converter G together,and distance transform loss is introduced to constrain the generation of edge information.2.To solve the problem that the road map after attribute conversion cannot be directly applied to the detection network due to the unlabeled information image,this paper uses a weak supervision method based on the teacher student network to solve this problem.The input pictures of the student network are scrambled after block coding,and the scrambled feature maps are restored to the relative order between the student network encoder and decoder before being sent to the decoder.The purpose of improving students’ network robustness is achieved through the operation of restoring the order after disrupting.To solve the problem that the network has weak ability to extract global information,this paper proposes pixel self attention to extract context information better.In view of the large loss of semantic information in global average pooling,a statistical feature channel module is proposed to replace the original average pooling to establish a global relationship.In terms of loss function,BCD(Bray Curtis Distance)and Hinge loss are introduced on the basis of cross entropy loss.By giving higher punishment to the points near the road edge that are judged incorrectly,the network has a better ability to fit the road edge.3.Based on the domestic expressway and urban road,this study has established a road detection data set that conforms to the actual road conditions in China.The data set we built includes 2649 daytime road pictures and 586 nighttime roads.And use Photoshop to calibrate the road picture.