Sift-based Scene Understanding Methods Study

Scene understanding based on image understanding, that is to say, you can understand the scene by processing scene image data. Scene understanding is high-level vision analysis. By processing and analyzing images or image sequences, we interpreted the contents of images and further more we gave the meaning of the scene. Generally speaking, any image understanding algorithm has a certain application conditions and limitations. Though there have many special image understanding systems, there has not a generic system in a broad sense of the word.This article has made the detailed introduction to image understanding, including image segmentation, feature extraction, feature representation and description, image matching and recognition. In order to correctly understand the contents of images, firstly, the object to be identified should be extracted from images. Then the target was recognized by using knowledge model. Finally, we got the interpretation of images. Image segmentation is to segment an image into some homogeneous regions, which are visually distinct and uniform with respect to certain properties. Image analysis and features extraction form a solid foundation for the implementation of image understanding. Effective features extraction is indispensable to image understanding.After the completion of image segmentation, feature extraction and description, the main issues to be addressed will be how to complete the interpretation of the image. Usually we can use the characteristic to describe the target in the course of image processing and understanding. This thesis includes the following contents:Several popular stable regions detectors and descriptors are discussed in details, and a new method based on maximally stable color regions(MSCR) is then proposed for scene recognition. Small maximally stable regions were merged or removed by using the method of mathematics morphology, which is more conductive to the following features extraction and description. By comparing our approach with the other two methods, results show that our method has a higher correct recognition rate. Also we compared our algorithm with a global appearance recognition method, and experiments demonstrate the effectiveness of our approach.In this paper, we exploit the MSCR detector to construct the maximally stable regions and the descriptor is computed using the scale invariant feature transform (SIFT), with the detected MSCR regions as input. The actual recognition of scenes goes as follows. First, we build a database with representative images for all the known scenes. Then we extract invariant regions and invariant feature vectors for all the images in the database, and store the feature vectors in the database. During the recognition phase, a query image is processed in exactly the same way. This image is matched by individually comparing each feature from the query image to this previous database and finding candidate matching features based on Euclidean distance of their feature vectors.The experimental results prove that this algorithm wins high recognition accuracy and robustness against non-linear image intensity transformation, a substantial range of affine distortion and changes in 3D viewpoint. Also we compare our algorithm to the global appearance based method, and show through experiments in both indoor and outdoor environments that our approach performs better.