Touchscreen devices, or even electronic stylus writing instruments (for example, Apple Pencils), are becoming increasingly common nowadays. Drawing sketches is something humans have done to communicate since ancient times. Sketches can be used to describe pictograms and allow the recognition of abstract objects and concepts; images that are similar to the sketches can also be retrieved through sketch-based image retrieval (SBIR).

The authors propose an approach to train and recognize similar sketches from a database, using both local features (for example, HOG, Daisy, SSIM) and holistic structure (star graph representation), and using machine learning techniques such as SVM and the multiple kernel learning (MKL) framework. The authors adopt a bag-of-features (BoF) representation and ensemble matching functions, and also propose to adopt/add “super-categories,” which can describe a set of sketches’ attributes (high-level attribute representation). These high-level attributes are effective in reducing confusion inside one super-category, and contribute to applications where the user enters the sketch and also a query describing some of these attributes.

The authors present the results obtained from several experiments using a well-known public sketches dataset, where they achieved significant accuracy improvements over other state-of-the-art sketch recognition methods.