Deep Stroke-Based Sketched Symbol Reconstruction and Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Hand-drawn objects are composed of meaningful parts.
• Accurate segmentation of sketched symbols is crucial for various applications.
• Existing methods struggle with the complexity of stroke-level symbol segmentation.

Purpose of the Study:

• To propose a novel neural network model for segmenting sketched symbols into stroke-level components.
• To introduce a reusable feature extractor for symbol segmentation.
• To provide a new, large-scale annotated dataset for the research community.

Main Methods:

• Utilized a fixed feature extractor based on the encoder of a stroke-recurrent neural network (RNN), a generative variational auto-encoder (VAE).
• Employed a multilayer perceptron (MLP) network to identify components based on extracted features.
• Developed a stroke-by-stroke symbol reconstruction approach.

Main Results:

• The proposed segmentation framework achieved superior performance compared to existing methods on a small dataset.
• A single encoder demonstrated reusability across multiple symbol categories with minimal accuracy loss.
• Extensive evaluations on a newly annotated large dataset confirmed significantly better accuracies than baseline models.

Conclusions:

• The proposed neural network model effectively segments sketched symbols into stroke-level components.
• The reusable feature extractor offers flexibility and efficiency for symbol segmentation tasks.
• The released dataset will facilitate future research in sketch-based symbol recognition.