Co-occurrence of Intensity and Gradient Features for Object Detection

For the purpose of object detection, Haar-Like Features (HLF) proposed by Viola [13][14] are very famous. To classify images, usually HLF and its extensions used only image intensity. However, it is well known that the gradient information of image intensity is very important for the object recognition [2][9]. So in this paper, we propose a feature which uses both intensity and gradient informations. Our feature, called “Co-Occurrence Feature (COF)”, can treat the co-occurrence of salient regions in both of intensity domain and gradient domain. We use an extended image set that consists of original (intensity) image and oriented gradient images which are extracted from original images. COF is composed from a pair of arbitrary rectangles on arbitrary image channel in the extended image set. As a result of face/nonface classification experiments, it is confirmed that our feature has good classification performance, especially in the high true positive rate zone of ROC curves, the false detection rate is significantly better than Viola’s HLF.