Prediction With Visual Evidence: Sketch Classification Explanation via Stroke-Level Attributions

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks excel at sketch classification but lack interpretable explanations.
• Existing visualization methods for sketch classifiers offer low-semantic regions and ignore inter-class correlations.

Purpose of the Study:

• To develop a novel explanation method for sketch classifiers that provides stroke-level evidence.
• To enhance the human understandability and discriminative power of sketch classifier predictions.

Main Methods:

• A sketch parser was developed to segment sketches into strokes while retaining geometric structures.
• A counterfactual map generator was designed to identify stroke-level principal components for categories.
• The method generates positive and negative semantic evidence to explain classification decisions.

Main Results:

• The proposed model effectively explains sketch classifier decisions using stroke-level evidence.
• Experiments on Sketchy-COCO and TU-Berlin datasets demonstrate the model's effectiveness.
• The method provides more discriminative and human-understandable explanations than existing approaches.

Conclusions:

• The novel explanation method significantly improves the interpretability of sketch classifiers.
• Stroke-level evidence and counterfactual analysis offer a more robust approach to understanding model predictions.
• This work addresses key limitations in current sketch classifier explanation techniques.