Interpreting Sloppy Stick Figures with Constraint-Based Subgraph Matching

Abstract

Machine systems for understanding hand-drawn sketches must reliably interpret common but sloppy curvilinear configurations. The task is commonly expressed as finding an image model in the image data, but few approaches exist for recognizing drawings with missing model parts and noisy data. In this paper, we propose a two-stage structural modeling approach that combines computer vision techniques with constraint-based recognition. The first stage produces a data graph through standard image analysis techniques augmented by rectification operations that account for common forms of drawing variability and noise. The second stage combines CLP(FD) with concurrent constraint programming for efficient and optimal matching of attributed model and data graphs. This approach offers considerable ease in stating model constraints and objectives, and also leads to an efficient algorithm that scales well with increasing image complexity.