基于图像仿射不变特征的多区域目标识别方法研究

Object recognition is a fundamental research problem in the area of computer vision. In a real-life scene, an object may be segmented into several sub-regions according to different attributes including color and texture. Due to the locality and non-uniqueness of sub-regions, object recognition methods which rely on the sole features of sub-regions have significant limitations on recognizing multi-region objects. If the viewpoints of images are sufficiently far, the distortion between images can be approximated by affine transformations. Consequently, this project aims at extracting affine-invariant features from the images of objects, analyzing the supporting and restriction relationship between features, and developing multi-feature fusion method for multi-region object recognition with high accuracy. By analyzing the variation of region distribution under affine transformations, the invariant feature of region distribution is developed. Based on these observations, affine region cutting methods of signal sub-region and sub-region pair are proposed, respectively. Then, the invariant geometrical and gray features are extracted. Through introducing the neural network and the supporting and constrain relationship between features, multiple features are fused to achieve a more stable sub-region matching probability matrix. Finally, a significance score is determined for each sub-region, and is used to measure the matching degree between the multi-region target and its corresponding model. The outcomes of this project will further boost the theoretical and technical development of computer vision in a number of real applications.

图像目标识别广泛应用于民用与军事的多个领域。实际应用中，不均匀的颜色、纹理等属性常导致目标被分割为多个子区域，而子区域的局部性与不唯一性令仅依靠子区域特征的目标识别方法存在很大的局限。为此，本项目拟在基于特征的识别算法中引入目标拓扑结构信息，以提高系统的可靠性与普适性。首先，为应对由传感器参数/成像视点不同引起的图像仿射变换，以挖掘仿射变换模型中的不变参数为基础，分别实现子区域与目标拓扑结构的仿射不变特征提取；继而引入神经网络与特征间的支持与约束关系，分别设计高效的类内特征与类间特征融合方案，提高子区域相似度对目标与模板匹配程度的衡量能力，并结合子区域的重要性评价，最终回答“如何度量多区域目标与模板匹配程度”这一核心问题。本项目旨在解决不受图像仿射变换及局部遮挡影响的多区域目标识别难题，其研究成果将为推动计算机视觉在目标识别、目标跟踪及末制导等相关方面的实用化进程提供重要的理论与技术支撑。

图像目标识别广泛应用于民用与军事的多个领域。实际应用中，不均匀的颜色、纹理等属性常导致目标被分割为多个子区域，而子区域的局部性与不唯一性令仅依靠子区域特征的目标识别方法存在很大的局限。为此，本项目在基于特征的识别算法中引入目标拓扑结构信息，以提高系统的可靠性与普适性。首先，以小孔成像模型为基础推导出目标形变被近似为仿射变换的合理性，并实现了区域分布参数变化建模及不变性分析；继而利用区域分布参数中的不变性，分别构建了描述区域分布的单区域特征以及描述区域间相对分布的区域拓扑结构特征，经过理论证明，上述两种特征具有严格的仿射不变性；在基于特征的目标识别阶段，为了提高识别精度，结合特征间的支持与约束关系，分别借助松弛标记与图表法实现了高效的特征融合，极大地提高了子区域相似度对目标与模板匹配程度的衡量能力。本项目的研究成果可望解决不受图像仿射变换及局部遮挡影响的多区域目标识别难题，将为推动计算机视觉在目标识别、目标跟踪及末制导等相关方面的实用化进程提供重要的理论与技术支撑。