Research On Outdoor Scene Image Segmentation Based On Convolutional Neural Network

Image segmentation is a crucial step in the field of computer vision,which provides rich visual information for subsequent correct recognition and detection by labeling different classes of objects.The development of fields such as unmanned driving and intelligent traffic analysis is closely related to outdoor scene image segmentation,which has an important role in daily life.At present,outdoor scene image segmentation mainly faces the problems of blurred segmentation edges and insufficient extracted feature information resulting in reduced accuracy of outdoor scene image segmentation.To address the above problems,two outdoor scene image segmentation models are proposed in the thesis.The main work is as follows:The outdoor scene images are affected by overlap,occlusion,motion and other factors,which make the segmentation results suffer from blurred segmentation edges.In order to solve the above problems,a network model SA-WNet is proposed,which combines squeeze and excitation blocks and attention gates.And adds the convolution operation to the U-Net to build a W-shaped network.Then,squeeze and excitation blocks and attention gates are introduced in the encoding and decoding parts respectively,which can make the network focus on the useful features and suppress the useless features.Finally,the combination of Dice loss and boundary loss enables the SA-WNet to focus on the global features while paying attention to the edge features.To address the problem that the traditional Seg Net does not extract feature information sufficiently,resulting in lower accuracy of outdoor scene image segmentation,an outdoor scene image segmentation network model Mobile-Seg Net is proposed.This network uses Mobile Net V2 network as the feature extraction part of Seg Net to obtain the feature information of outdoor scene images.And the pyramid pooling module is fused after the encoding part to extract features at different scales and obtain global context information,which enables the Mobile-Seg Net to fully extract the feature information in outdoor scene images.In the thesis,the proposed network model is experimented on public dataset.The experiments show that the SA-WNet reached 0.9198,0.7502,and 0.6412 in pixel accuracy,mean pixel accuracy,and mean intersection over union,respectively,which is better than most of the comparison network models;the Mobile-Seg Net has improved pixel accuracy and intersection over union for each category in the images compared with the Seg Net.