Research On Urban Pavement Image Segmentation Based On Feature Self-interference Mechanism And Attention Mechanism

The research of urban road segmentation is to use the computer vision technology to divide the road information lines in front of the vehicle for the recognition of road obstacles.The difficulty lies in the fact that the road surface detection is easily affected by the weather,such as light,shadow,and so on.At present,semantic segmentation is an important technology in the area of computer vision,which is widely used in the task of urban pavement detection.The mainstream pavement detection methods include semantic segmentation-based full convolution network(FCN),UNet,Deep Lab,etc..Their common convolutional neural network is the supervised learning method,which can automatically learn the features of the image,end-to-end classification learning to achieve pixel-level classification or segmentation.However,as the depth of convolutional network increases,the network parameters increase rapidly,so it is easy to over-fit.Therefore,anti-interference training is needed to improve the generalization ability and robustness of network features,and to reduce over-fit.Another way to improve the robustness of features is the attention mechanism,such as the use of Non-local Algorithms to fully calculate the global relevance of feature graphs to play the role of feature optimization,but the lack of channel feature screening.Based on the above considerations,we propose the feature self-interference mechanism in the process of anti-jamming and the Pixel attention mechanism in the improvement of the attention mechanism.The main research work of this paper is as follows:1.In order to solve the problem of lacking anti-jamming ability and the difference between the interference sources and the original features,this paper proposes a self-jamming mechanism,which adds the feature self-jamming network to the existing main segmentation network,by using the original image clipping and interpolating,the pseudo-feature image is sent into the self-interference network to generate the interference information,and the interference information is combined with the original information through the feature fusion module,in order to improve the robustness of the model,the main segmentation network needs to overcome the interference and segment the target accurately.A feature fusion module is designed between the backbone network and the jamming network to realize the self-interference of the pseudo-feature to the real feature.2.In the Non-local attention mechanism,the input feature map is too large,which results in too much matrix transposition computation,while the channel attention mechanism has too few training parameters and lacks pixel-level feature selection.In this paper,the attention mechanism of pixels is proposed,and the attention information between pixels is correlated by means of full connection and convolution,and the feature pixel level is filtered by means of Sigmoid modulation,it reduces GPU memory consumption and increases trainable parameters.A multi-scale feature extraction module is introduced in the last layer of the main network encoder,and four different-size convolution cores are used to extract the image features to expand the receptive field and improve the fine granularity of the segmentation network.3.Considering the influence of the size of data set on the normalization effect,this paper adopts the combination of group Normalization(GN)and batch normalization(BN),the main network and the auxiliary network adopt group normalization and batch normalization to avoid the influence of training batch reduction on the network precision.The experimental results show that the proposed method can improve the segmentation accuracy from 88.02% to 90.55% on Kitti and Cityscapes road database compared with Deeplab V3 segmentation method,the segmentation accuracy on the Cityscapes dataset was improved from 87.15% to 90.16%.