Semantic Segmentation Based On RGB And Depth Information

Image semantic segmentation refers to dividing the image into blocks with semantic labels according to words information.It is the key step of image analysis.It can be applied in robot navigation,automatic driving,multimedia technology and other fields.It can effectively improve the image segmentation with the RGB information and deep image information.In recent years,it has achieved great success in the field of semantic segmentation for indoor scenes.Due to the complexity of indoor scenes and the diversity of the layout,it is still very difficult to obtain accurate indoor scenes.Therefore,this paper proposes an image semantic segmentation algorithm based on RGB and deep image information.The main research work is as following:1)First,it is about object segmentation algorithm of indoor scene based on RGB-D images.This algorithm is mainly put forward because of difficulty of excessive segment in foremost segment.On the basis of the initial segmentation,we execute the perceptual grouping.Through further extraction of the characteristic of super pixels it combines the similar super pixels and it is effective to correct the problems of the image segment.At the same time,color information is vulnerable to the effects of external factors such as illumination and environment.If the colors of two adjacent objects in the image are too close,it is very difficult to tell them from each other.In this algorithm,the color information and deep image information are used at the same time.Experiments prove that if we combine the two kinds of information effectively it will help to improve the precision of object segmentation and enhance the precision of object segmentation as a whole.2)Second,it is about semantic annotation algorithm based on multi features of indoor scenes.It is crucial to put the semantic labels of objects into the corresponding segmentation blocks.In order to obtain the classifier with high accuracy,we try to extract region's color,shape and position(direction,height)in the region feature extraction phase.Because of the variety of indoor scene objects and uneven distribution of samples,the random forest with strong generalization ability is used as classifier to train and test.Experiments prove the effectiveness of the proposed algorithm.