Research On Scene Image Classification Application Based On The Robot Vision

Scene classification is a hot and important research direction in computer vision field. It is meaningful for people to study how human understand the semantic meaning of images. Recently years, scene classification has been a practical and important way in those fields like image retrieval, medical image recognition, computer vision and robotic. With the increasing digital images, it is becoming harder to classify the images with the classical way that label images manually. Thus, it becomes more and more important to find a much more intelligently way to classify images. The process of scene classification includes extracting features of images, building visual dictionary, predicting the classified results. When builds a visual dictionary, a reasonable capacity of the dictionary is the key point which will affects the accuracy and efficiency. In allusion to the way which gets the capacity by repetitive human works, invite the AP clustering algorithm. This way can not only get the capacity automatically which means avoid those human works, but also lift the efficiency of the algorithm greatly. The main content of this paper as follows:Firstly, this paper introduced the background and research significance of scene classification. It also analyses the current situation of the research of clustering algorithm and visual dictionary.Secondly, this paper shows the flow chart of scene classification. And states the process of get SIFT and CTH feature, and AP clustering. Then it analyses the advantage and disadvantage of the 2 clustering ways by compare the randomly producing 100 data with the clustering result. The number of clustering center of K-means clustering must designate by human previously, and the different initial clustering center will cause the different clustering results. While, AP clustering obtains clustering center automatic by iterative algorithm, and in the process of iteration, the result will become stable after 44 times. Furthermore, this paper also analysis the 2 parameters of AP clustering how to impact the clustering result. The experiment shows that AP clustering is better than K-means when there are no experiences of related research.Thirdly, this paper make a further study to the way of extract the SIFT feature and CTH feature, and provide several ways to build the visual dictionary by using SIFT feature. Then compare the grid sampling with random sampling through experiment simulation. The result shows that grid sampling is much more suitable to the scene classification in this paper.Fourthly, this paper preprocesses the extracted characteristic matrix data. Then it obtained the capacity of the visual dictionary with the K-means clustering algorithm and AP clustering algorithm respectively, and uses SVM to do the scene classification. Find a relation curve between the visual dictionary capacity and recognition rate using K-means clustering. And then changes the parameters of AP clustering algorithm, analyses the effect of the parameter on the scene classification. Finally, it evaluates the way to classify clustering results by IGP.Fifthly, this paper analyses the results of the experiment. Firstly, it compared the relation curve between recognition rate and visual dictionary capacity obtained by K-means clustering algorithm with obtaining visual dictionary capacity automatically by AP clustering algorithm. And find that AP clustering algorithm can obtain visual dictionary capacity more easily, and the recognition rate of the scene classification can be over 81%. While when using the K-means clustering algorithm there are no relation curve between recognition rate and visual dictionary capacity. And the highest precision of the classifications can only be 350,750, 1100 and 1350. What is more, it also shows that several parameters can affect the results and how they affect it. Besides, it verifies the effectiveness of AP clustering algorithm and the rationality of the obtained capacity of the dictionary. Finally, compared the SIFT feature description and CTH feature vector to describe the two methods of feature extraction and classification of scene image, the different effect of CTH has been proved by the data characteristics can be more accurately and quickly realize image scene classification.Sixthly, this paper summarizes those works done in experiment. And give some advices to the further study.