A Method For The Scene Classification And Labeling Based On Multi-feature Fusion

Scene classification is a new research field and it rised in the late 1990 s. In recent years it has become the hotspot research.As the multi-disciplinary cross technology, such as computer vision, artificial intelligence, cognitive science, database, pattern recognition and human-computer interaction and so on, scene classification is widely attention. It can be used for image retrieval, mobile robot systems, intelligent video surveillance, medical applications, and military applications, etc., and the main research of scene classification is how to use the computer to segment, recognize, and classify images or videos quickly, accurately and effectively. And it make computer classify images into the specific semantic types according to the semantic content of images, using the method of unsupervised.In this paper, first of all, it summarizes the current research status of scene classification at home and abroad. Then a method for the scene classification and labeling based on multi-feature fusion is proposed. This paper mainly divides into three parts:1. In global image scene classification and labeling method, it introduces the method of image median filtering, image segmentation algorithm, image feature extraction and fusion method and principle of support vector machine classification. Firstly, the median filter is used to suppress random noise and at the same time don't make the edge blur, which is good to preserve the image edge information characteristic of image filter processing. And then we use the Mean-shift algorithm for image segmentation, after that we must extract the contour of the image.Through setting the area and perimeter's threshold of outline area, we can merge adjacent areas of the split image, and fill the outline of the combined areas. Through this method we can eliminate over-segmentation phenomena appeared in the image further and we can achieve a good result for image segmentation. Finally, after extracting color features and the texture features of those segmented regions, the feature can be fused which would be input to the SVM classifier. To sum up, we have completed the global image scene classification and labeling.2. In view of the shortcomings of the traditional method which is based on visible image and object modeling for the object classification and labeling, such as time-consuming and accuracy not enough, this paper puts forward a new kind of method of object classification and labeling based on visible light and laser radar image fusion, and it make full use of the rich texture information of visible light image and the rich height information from the laser radar image classifying and annotating object. Firstly, we achieved the general location area coordinate information of the object in the visible image using the target's height information in laser radar image. Only for these areas it is greatly shortened the time to find the target area and reduce the error detection rate. Then we must extract the visible light image area to the location of coordinate and the corresponding optical image area. After extracting SURF_BOF features of those areas, we put them to SVM classifier for the category classification. Finally we finish the object classification and labeling.3. According to the proposed image global scene classification and labeling method, visible light and laser radar image fusion of object classification and labeling method, we realize the scene classification and labeling software which based on multiple feature fusion. The software can achieve objects classification and marking in the global image scene, the classification and labeling of object. At the end of the software it has finished the classification and highlight of the obstacle when we simulate the flight of unmanned aerial vehicle.