Research On Scene Recognition Based On Local Feature Matching

Scene recognition is an important research content of computer vision and plays an important role in applications such as visual navigation,visual information retrieval,image and video understanding,and has received increasing attention in recent years.The characteristics of scene recognition and its technical difficulties are analyzed in this paper.From the aspects of feature extraction,feature matching,and geometric matching,the scene recognition method based on local feature matching is studied.The common feature extraction methods were studied in this parer,and the advantages and disadvantages of SIFT features and SURF features were compared and analyzed.For the stability and rapidity requirements of local feature extraction,the SURF-SIFT features that combine the advantages of SURF and SIFT are proposed.The real scene image is used for feature extraction and matching experiments.The results show that the SURF-SIFT feature is superior to the SURF feature and SIFT feature in detection speed and matching accuracy.A scene recognition method based on local feature matching is introduced,which uses feature clustering,inverted index and KD tree to realize fast scene recognition.For the clustering speed problem of commonly used clustering methods,the method of picture-by-picture clustering is adopted to speed up.To solve the problem of slow search speed of high-dimensional features in KD-tree,BBF fast retrieval is used to speed up matching.Experiments were conducted on the architectural scenes of the two public datasets of Oxford Buildings and 201 Books and CTurin180.The results show that this scene recognition method has advantages in speed.For the situation of error scene in the returned result for scene recognition,the image geometric matching method was used to filter the recognition results.The principle of the RANSAC method is introduced.Then,aiming at the shortcomings of RANSAC,the spatially consistent RANSAC is introduced.Spatial relationship checking is used to eliminate some of the incorrect match pairs,and then RANSAC is used for purification.In order to extract the target in the scene,a foreground extraction method of the shape model is proposed learning from the ISM.Scene recognition experiments are performed on the scenes of two public data sets.The results show that the scene recognition method in this paper has advantages in speed and accuracy.