Copesetty, Siddhartha Karthik (2016-08). Integrating Multiple Sketch Recognition Methods to Improve Accuracy and Speed. Master's Thesis. Thesis uri icon

abstract

  • Sketch recognition is the computer understanding of hand drawn diagrams. Recognizing sketches instantaneously is necessary to build beautiful interfaces with real time feedback. There are various techniques to quickly recognize sketches into ten or twenty classes. However for much larger datasets of sketches from a large number of classes, these existing techniques can take an extended period of time to accurately classify an incoming sketch and require significant computational overhead. Thus, to make classification of large datasets feasible, we propose using multiple stages of recognition. In the initial stage, gesture-based feature values are calculated and the trained model is used to classify the incoming sketch. Sketches with an accuracy less than a threshold value, go through a second stage of geometric recognition techniques. In the second geometric stage, the sketch is segmented, and sent to shape-specific recognizers. The sketches are matched against predefined shape descriptions, and confidence values are calculated. The system outputs a list of classes that the sketch could be classified as, along with the accuracy, and precision for each sketch. This process both significantly reduces the time taken to classify such huge datasets of sketches, and increases both the accuracy and precision of the recognition.
  • Sketch recognition is the computer understanding of hand drawn diagrams. Recognizing sketches instantaneously is necessary to build beautiful interfaces with real time feedback. There are various techniques to quickly recognize sketches into ten or twenty classes. However for much larger datasets of sketches from a large number of classes, these existing techniques can take an extended period of time to accurately classify an incoming sketch and require significant computational overhead. Thus, to make classification of large datasets feasible, we propose using multiple stages of recognition.

    In the initial stage, gesture-based feature values are calculated and the trained model is used to classify the incoming sketch. Sketches with an accuracy less than a threshold value, go through a second stage of geometric recognition techniques. In the second geometric stage, the sketch is segmented, and sent to shape-specific recognizers. The sketches are matched against predefined shape descriptions, and confidence values are calculated. The system outputs a list of classes that the sketch could be classified as, along with the accuracy, and precision for each sketch. This process both significantly reduces the time taken to classify such huge datasets of sketches, and increases both the accuracy and precision of the recognition.

publication date

  • August 2016