Natural Scenes In The Identification Of Common Landscape Research

Scenery recognition system, which this paper researches as a kind of image recognition system, can recognize common objects in the environment, which is the object of the sub project from National Defense Department. This system belongs to the area of image processing, which is a part of pattern recognition and artificial intelligence. It bases on the ATRV-2 mobile robot platform, and can process, and recognize images shot by the robot. These images are divided into two average halves, the top half is recognized as a whole object, and recurring to another sub project of the whole project, the lanes in the bottom have been found out, so the part between the lanes and the part outside the lanes in the bottom half are recognized respectively. At last, common objects such as human beings, cars, buses, buildings, cats, bicycles, and so on can be recognized.First of all, this paper concerns about several kinds of general features, such as color histogram, color moment, wavelet texture, shape invariable moment, and so on, so as to choose some of them, which are more efficient to this system and being used more widely. Then, SVM (Support Vector Machine) is used to train images of 12 kinds. In the process of training, different combinations of those features, different SVM parameters, and different SVM kernel functions are used to decide the effects of them on the recognition results. At last, SVM is still used to predict which kind the input images belong to. The result shows that some features can really improve the recognition results, but others are not, and some even make the result worse. Also, the parameters and kernel functions of SVM can affect the result, too.After testing on 834 images, which is either taken by the robots, or retrieving from the Internet, it is proved that, using some features, parameters and kernel functions, the system can achieve good results. In the mean time, this system can be extended, that is, more features can be added, and trained by SVM to improve the recognition effect. And if more images are used as training samples, the result can be improved, too.