Application Research Of Image Segmentation Algorithm Based On Deep Learning

With the continuous development of artificial intelligence technology,computer vision has achieved remarkable results in image processing.Under the influence of deep learning,image segmentation has experienced an important transition from traditional image processing techniques to deep learning dominated ones.Now more and more applications require inferring relevant semantic information from images for subsequent processing,such as autonomous driving,video surveillance,augmented reality,etc.Therefore,this thesis systematically studies the image semantic segmentation algorithm based on existing deep learning methods,and summarizes the content as follows:Firstly,this thesis proposes an improved complex scene image semantic segmentation algorithm based on PSPNet,called ESD-PSPNet.It solves the problem of inaccurate semantic recognition of some targets that are too small or easily overlap with other targets in some complex scenes.This thesis improves PSPNet network by combining superpixels,and replaces the original bilinear interpolation upsampling with dense upsampling in the pyramid pooling module.Finally,the dataset is preprocessed to make the image effect better after semantic segmentation.The experimental results show that ESD-PSPNet semantic segmentation model has good performance for small object segmentation and meets human visual requirements.Secondly,this thesis proposes a road image semantic segmentation algorithm based on DeeplabV3+ improvement,called Deep SENet.The algorithm solves the problem of insufficient utilization of edge feature information in the original network model.It also solves the problem of lack of dependency among feature information,which leads to local information loss and inconsistency of semantics and scale among input feature information.This thesis adds an edge information processing module to DeeplabV3+ network to enhance the semantic representation of features.At the same time,in order to solve the problem of lack of dependency among information in the atrous spatial pyramid module,which leads to local information loss,this thesis replaces the original network with a channel multi-scale perception module.This makes the network focus on the required feature channels and ensures the globality of the network.Finally,this thesis uses a multi-scale attention feature fusion mechanism to replace the original network’s simple feature fusion method,which improves the performance of semantic segmentation.The experimental results show that Deep SENet semantic segmentation model has obvious performance improvement on Cityscapes dataset.Finally,this thesis designs an image semantic segmentation system for autonomous driving scenarios,based on the semantic segmentation algorithm proposed in the previous section.It solves the problem of poor environmental perception ability of current autonomous driving technology,which affects the vehicle’s ability to make safer driving decisions.The system consists of image acquisition module,image preprocessing module,image clarity detection module and image semantic segmentation module,which can achieve high-quality semantic segmentation.The system can be applied to autonomous driving field and provide help for developing smart city and intelligent transportation system.