基于空间位置关系改进的物体识别上下文模型

A Context Model for Object Recognition Improved by Spatial Relationships

常见的物体识别算法是基于图像局部低维特征的,在图像成像质量较差、分辨率较低情况下存在不确定和歧义性;图像上下文包含了场景信息以及物体之间彼此关联的丰富信息,可作为图像低维特征补充从而有助于提高物体识别率。该文在已有上下文模型基础上,进一步考虑了物体的空间位置关系信息,将图像全局特征、物体同现性关系和空间位置关系信息、局部检测器输出整合到同一个概率框架中,并充分利用树结构图模型高效推理的优势,改进了物体识别性能。最后通过标准图像集进行算法验证测试和对比来说明该文算法的有效性。

Most object recognition algorithms are based on the local low-dimensional characteristics of the images.Images of poor quality or low resolution prompt uncertainty and ambiguity for object recognition .Context information contains scenes and spatial relationships information among objects which can be used as the supplementary for low-dimensional image features to improve the recognition .In this paper , we present an improved context model for object recognition in which we further consider the spatial relationships .We integrate all the information in a unified probabilistic framework taking full advantage of the efficient inference algorithms of tree models .Our model can improve the recognition result and give a consistent scene interpretation of the image.At last, we use a standard dataset to test our algorithm and make comparison to show the effectiveness .