İnsanların özenli çizim algısının öğrenilmesi

The increasing availability of pen-based tablets and pen-based interfaces resulted in several lines of work that attempt to build computational tools for processing sketches. These include computational methods for classifying sketches and methods for beautifying rough drawings. In this thesis, we pose discrimination of clean sketches from messy ones as a new problem, and propose a computational framework for learning messiness from examples. Our system does not use domain, object or context specific knowledge. Our approach is inspired by human perception and can assess the messiness of hand-drawn sketches using features extracted from relative orientation, length and distance properties of primitives. In order to evaluate our method, we collected a sketch dataset that consists of 1535 drawings of 32 different symbols, which were also coded with subjective messiness information obtained from human raters. Our evaluation shows that we can discriminate messy and clean drawings with up to 90% accuracy when the problem is treated as a 2-class classification problem. We report detailed analysis on what makes this problem hard, and how the system accuracy relates to the inter-rater agreement on the messiness ratings of sketches.