Sketch Classification with Convolutional Neural Networks

A deep learning approach to classifying hand-drawn sketches across 250 object categories, achieving 58.7% top-k accuracy on the TU-Berlin dataset using a custom CNN architecture.

Abstract

We present a deep learning model to recognize hand-drawn sketches. This is an image recognition task and part of multi-class classification using deep learning. We show an overall 58.7% top-k categorical accuracy and that the dataset is well suited for image classification using deep learning. Furthermore, we show that tuning the model parameters -- i.e., layers, filter size, and preprocessing -- significantly impacts the results.

Key Contributions

  • Designed a 14-layer CNN architecture (8.6M parameters) for 250-class sketch classification on the TU-Berlin dataset of 20,000 hand-drawn greyscale images
  • Demonstrated that top-k categorical accuracy is the appropriate metric for hand-drawn sketch classification, where visual ambiguity across classes is inherent
  • Showed that image scaling (128x128), normalization, and grayscale preprocessing significantly impact model convergence and accuracy
  • Achieved 58.7% top-k test accuracy in just 5 training epochs, outperforming earlier SVM-based approaches (56%) on the same dataset
  • Explored the effects of hyperparameter tuning using AutoKeras and KerasTuner, settling on a sequential model with progressively deeper convolutional blocks
  • PDF: to be hosted
  • arXiv / TechRxiv: to be added

Back to Research & Papers