Scene Classification Based On Bag-of-visual-words And Sparse Representation

The first problem to be solved should be the robot localization for robotics research, but thetraditional self-positioning and navigation technology is only able to achieve that robotcognitive geographically environment and does not recognize the surrounding environmentsemantic information. Because traditional environment map is the lack of semanticinformation, resulting in that robot cannot recognize the surroundings. To solve this problemSouth China University of intelligent software and Robotics Laboratory Dr. Liang Mingjieproposed the idea to create a semantic map, which is based on the traditional map addedsemantic information of the surrounding environment and is used to describe the environment.The first thing to be achieved is to make the robot can recognize the surrounding scene, whichis,in the field of computer vision, to study the subject of the image scene classification.SecenRecognition is the basic problem for creating a semantic map, therefore, it is the first problemto be solved for achieving recognition scene, that it is image scene classification problems.Based on previous research on the image scene classification, it founds that two majorproblems still exist in the most of the classification algorithms:(1) They do not support wellmultiple image scene categories and with the number of classes increasing, the effect ofimage scene classification algorithm significantly decreased;(2) Most classificationalgorithms can not cope well in the image blurring, mutilated and noise and other factors,resulting in relatively poor image scene classification. Words package model can create alocal semantic concept image, describing high-level semantic information of image,and thus itcan be better able to describe and classify images. Sparse representation in face recognition,image noise removal, blind source separation, DOA estimation and image restoration researchareas has a very important significance, and for the image noise, fuzzy, incomplete, and otherfactors have relatively good treatment effect.Firstly, the Bag-of-Visual-Words model process includes the extraction of key points,construct and building a dictionary word wrap, then we use KNN and SVM classifier toclassify a test image and analyze the experiment results, and experiments show that the bow model has a better effect on multiple scenes. Secondly, we use statistics of the imagehistogram with KNN, SVM and sparse representation classifier respectively for imageclassification, then the experimental results were analyzed and SRC (sparse representationclassification) classifier has been proved that it has better classification results.Secondly, the statistics of the image histogram, then use KNN, SVM and sparserepresentation classifier (Sparse Representation SRC) for image classification, theexperimental results were analyzed, SRC classifier has proved better classification results.Finally, the image scene classification based on Bag-of-Visual-Words and sparserepresentation algorithm is described based on the above two experiments. Then, theexperimental results were analyzed. Experimental results show that it has a better effect onclassification than the simple use of the word wrap and sparse representation, and accordingto the text classification ideological, the algorithm can be expanded well and optimized, so ithas strong application value.