Semantic Image Segmentation Based On Superpixel Segmentation And Graph Neural Network

Today,artificial inteligence technology has influenced kinds of aspects in people's life.While,some problems in artificial inteligence processed by deep learning algorithm,such as computer vision methods and natural language processing methods.Among these problems,the application which has the biggest influence in ordinary being's daily life,may be the computer vision.Because of face recognition has been applied in opening the door and passing in rail way station,people begin to realize the power of this new technology.And like manless driving and the city brain,these new concepts appear in real life.These new technologies bring about different ideas of life.So computer vision is processed in deep learning method,attract more and more researchers' attention.Even so,there are lots of unresolved problems remaining.Under this trend of the times,the author of this paper do some research work in image segmentation.Image segmentation has three main research directions.They are image se-mantic segmentation,image instance segmentation,and panoptic segmentation.This paper is about the image semantic segmentation.Due to the author of this paper has observed some appearance of recent algorithms'shortcoming.Just like some boundary message lost,and some noise blocks were in segmentation result picture.Aim at processing these two re-maining problems,we investigate some different papers and do logical experiments to propose a new algorithm.This image semantic segmentation algorithm can be divided into five parts,like basic segmentation module,superipixel segmentation module,feature extracting module,graph neural network module,and the decoder module.Basic segmentation module is mainly composed of deeplab v3 model,and different segmentation models could be used in this module.The experiments proved that our algorithm has effect on different basic segmen-tation models.We use SLIC algorithm in superpixel segmentation module.About feature extracting module,we use LAB color model and convolutional neural network to extract low-level and high-level different layer's feature.About graph neural network module,we choose graph attention network as a regular item for basic segmentation network.About decoder module,we choose this module design to keep the low-level feature.Besides proposing an image semantic segmentation model,which is based on superpixel segmentation and graph attention network.This paper aims at clarifying the details'design,and hyper-parameters.So we carry through a series of experiments.And get the training and test result in Pascal Context dataset and Pascal VOC 2012 dataset.At the same time,the direction of this paper's work has lots to do.In future,our method can be used in doing research in panoptic segmentation.With the help of graph nerual network,some interesting works could be produced in image segmentation.