A stochastic hierarchical model and nonparametric statistics approach to object recognition
Antonio Zelic · 1996
We propose a new approach to multiple object recognition problem. The approach combines and explores conceptual ingredients from formal grammars, Markov Random Fields with polygonal realizations, and nonparametric statistics. It contains two major models: (a) Stochastic Hierarchical/Syntactic Models representing the overall shape architecture, and (b) data models describing the interaction between the SHM and the image data. A key property of our method is that there is no preliminary segmentation; in fact, segmentation and recognition are simultaneous processes. The data models are designed so that object recognition and scene interpretation are robust (invariant) with respect to lighting, contrast and degradation effects; this is attained by employing nonparametric statistics which are invariant under monotone transformations. The SHM has multiple levels of hierarchy and syntax. The top levels view objects as concatenations of their articulated parts, and are represented by a directed graph structure (membership graph). The bottom level views boundaries of top-level primitives as concatenations of low-level elementary units; this concatenation process may be represented as a Markov process with dynamic evolution of the set of allowable states. The approach leads to a formulation of object recognition as a global optimization problem which, in view of its recursive structure, lends itself to variations of dynamic programming. The state space explosion prohibits the possibility of exact computations, and hence efficient search strategies, initialization, pruning and backtracking techniques are devised. Our approach has been successfully tested on a database of 2-D manual tools in environments highly degraded by noise, blur, clutter and occlusion.