Scene Classification Based On Multi-feature Extraction

Scene classification is an important branch of machine vision,which provides certain priori information for behavior recognition,target detection and other fields.It is widely used in large database image and video retrieval,intelligent robot and intelligent security.Because of the complexity of scene images,the similarity between classes and the confusion within classes are still a challenge problem to be solved.In this thesis,the research on RGB-D indoor scene recognition and natural scene recognition are carried out.An indoor RGB-D scene classification algorithm based on improved SIFT feature and depth network and a natural scene classification algorithm based on local feature and global feature fusion network based on ROI are proposed.The main work of this thesis is as follows:(1)An indoor RGB-D scene classification algorithm based on improved SIFT feature and depth network is proposed in that the improved SIFT feature,high-level semantic feature and depth histogram feature are extracted.The improved SIFT feature is the SIFT feature with little edge effect selected by inputting the weighted Euclidean distance between the SIFT feature point and the Canny edge point into random forest,and the high-level semantic feature is extracted by using the convolution neural network based on Res Net.The depth histogram feature is a feature which describes the depth information of the image by using the histogram.The improved SIFT features,high-level semantic features and weighted depth histogram features are concatenated to get the RGB-D scene features,and the Soft Max classifier is used to classify them.(2)A natural scene classification algorithm based on the fusion network of local features and global features based on ROI is proposed.Firstly,the region of interest is sorted according to the contribution degree of the object category of the region of interest to the scene classification,and then the region of interest which has great contribution to the classification of the scene is selected to extract the local features around the region of interest.The local feature is weighted according to the contribution degree and fused with the global feature.Finally,the Soft Max classifier is used to classify the natural scene with the input of the original graph into the BN_Inception-based dual network model.In this thesis,the proposed indoor RGB-D scene classification algorithm and natural scene classification algorithm are tested on NYUD v2 dataset,MIT-67 dataset and Scene-15 dataset,respectively.The results show that compared with the current mainstream classification algorithms based on manual features and deep learning,the recognition rates of the two algorithms proposed in this thesis are improved to a certain extent,and the two algorithms have good robustness.To some extent,the edge effect of SIFT feature,the low level feature of convolutional neural network missing image and the randomness of local feature extraction are solved to some extent.