Agent-based sketch recognition

A technique for the classification and ‘understanding’ of ‘child-like’ sketches is explained. Using a live pen-based input device, sketches of animals are classified using our system: SIVA (Sketch Understanding Via Agents). We separate the task into three sub-problems: Sketch segmentation, segment matching and a contextual conflict-resolution system. A successful sketch segmentation method involving corner detection using the perpendicular distance from a chord to the line in addition to pen speed is shown and evaluated. A grammatical curve description is then used to match fragments of the sketch to model fragment descriptions in a database. Finally, we discuss a method to verify the identified fragments and recognise the subject of the sketch by examining the compatibility between the fragments found in the sketch. We achieve this using an agent-based search of the sketch – agents representing particular features are ‘released onto’ the sketch and interact with one-another and sketch fragments, and in attempting to achieve their own goals achieve recognition of the sketch.