| Image semantic segmentation technology is one of the important means used in the field of unmanned vehicle autonomous driving.However,at present,unmanned vehicle autonomous driving still has the problems of difficult segmentation effect,difficult data set production and excessive cost.To solve the above problems,this paper designs a multi-channel convolution residual structure MCRB,two semantic segmentation networks MCRB-Net and AMCR-Net in the semantic segmentation structure design,and verifies their effectiveness in improving the segmentation effect through ablation experiments.Then,two data augmentation methods,A-Pix2pix and M-Grid Mask,are designed to augment the cityscapes dataset.Finally,the effectiveness of the two data augmentation methods designed in this paper is verified through comparative analysis.The specific research contents are as follows:（1）Aiming at the problem that it is difficult to extract sufficient feature information from conventional convolution structure,a multi-channel residual convolution structure MCRB is designed from the aspects of residual structure,multi-scale feature fusion and multi-channel convolution.According to this structure,a new semantic segmentation network MCRB-Net is designed.In order to capture the dependencies between pixels and channels,a dual attention mechanism module is introduced.Finally,a new semantic segmentation network model AMCR-Net is designed by using the multi-channel residual convolution structure MCRB and the double attention mechanism module.The ablation experiment verifies the effectiveness of the network structure designed in this paper.（2）Two kinds of data augmentation methods are designed to solve the problem that it is difficult to collect and make training data sets in the field of automatic driving.The first data augmentation method is based on conditional generation of countermeasure network pix2pix.Aiming at the problem that its generator is difficult to generate fine samples,this paper uses AMCR-Net to replace the generator,and redesigns A-pix2pix method.Aiming at the problem that it is difficult to collect occlusion samples in real scenes,this paper improves the mask image generation rules and designs an M-Grid Mask method based on Grid Mask,which is an information deletion data augmentation method.In order to effectively control the deletion and retention of the edge area of the augmentation data,the m-gridmask method is proposed.（3）On the basis of AMCR-Net and U-Net,this paper conducts experiments based on two data augmentation methods,evaluates the two data augmentation methods designed in this paper through the comparative analysis of semantic segmentation effects,and verifies the effectiveness of A-pix2pix and M-Grid Mask. |
