Image Middle-level Description Based Scene Classification Research

With the rapid development of computer vision technology,the requirements for image understanding are also increasing.Scene classification as a branch of image understanding direction has very important research significance.This paper studies the image scene classification by forming a more effective image description,aims to train a more discriminating scene classifier and improve the recognition rate of image scene classification.First of all,this paper established a spatial mesh classification model for image multi-scale description,and added scale factors to the spatial pyramid model to transform the three-level space pyramid grid of the same scale into a three spatial grids of different scales.An improved K-means algorithm was proposed.Not only does it effectively avoid local optima,but it also automatically selects the right number of cluster centers based on the structural characteristics of the data.At the same time,rationality evaluation was performed on two data sets with different sizes.Experiments show that multi-scale image description methods achieve better results in classification.Secondly,the method combined the local and global features of the scene image by finding differences and complementarities between different feature descriptors,integrated global features and local features effectively.At the same time,it analyzes and compares the performance of three different fusion methods based on feature description,low-level features,and spatial pyramids.At last,the global features are added on the basis of two different local features to form a more effective image description.Finally,this paper proposed a space strengthening grid algorithm.This algorithm performs prior knowledge intervention on the parameters of the meshing,and at the same time enhances the space of the semantically rich area,making the image description more focused on the distinct semantic regions,and thus forming a more effective image description.The low-level features and high-level semantic features are both used to classify the model.And finally,it compared with the general spatial grid,which verifies the effectiveness of the algorithm.