| （英） |
This paper presents a new method for improving classification performance based on the SVDD including a target object detection scheme. In the proposed method, regions including the target objects are automatically selected based on the distance between their visual feature vectors and the center of the hypersphere calculated by the SVDD. Then a generation of new positive examples is realized. Thus, even though from training images with various position, direction, scale, and shape of target objects, it can be expected to use only local blocks including the target objects as the new positive examples. Furthermore, by using also local blocks including the selected regions, it can be realized to increase a variation of the positive examples including the target objects. Therefore, the automatic selection of the regions including the target objects and the generation of the new training images based on the obtained regions become possible, and a highly accurate classifier is realized. Experimental results are shown to verify the performance of the proposed method. |