Abstract
Generating realistic free-hand sketches is a significant challenge due to the sparse, geometric nature of the medium. We present a novel Generative Adversarial Network (GAN) framework that disentangles content (what is drawn) from style (how it is drawn). By modeling style as a latent geometric variable rather than a textural one, our approach generates diverse, high-quality synthetic sketches. Experimental results on the SketchIME dataset demonstrate superior performance in realism (FID 12.175) and content accuracy (>96% for unseen classes), offering a robust solution for data scarcity in specialized domains like military symbology.
Introduction
In specialized fields such as military situational mapping, free-hand sketches are essential for communication but difficult to collect at scale. Unlike common objects such as animals or everyday items, hand-drawn military symbols lack large-scale datasets, which hinders the development of automated recognition and segmentation systems.
The primary difficulty in synthetic sketch generation lies in the definition of "style." In sketches, style is not a textural property but is embedded in geometric nuances: line thickness, slant, and structural imperfections. Standard image-to-image translation models, which typically rely on texture-based style transfer, often fail to capture these structural characteristics. Furthermore, individual artists exhibit unique "painter styles," adding another layer of complexity to the generation task.
Methodology
The proposed framework utilizes a Content-Conditioned GAN architecture designed to handle the sparse nature of sketches. The system is built around four primary components:
- Generator (G): A U-Net-based architecture with skip connections that transforms a standardized "content icon" into a stylized sketch.
- Style Encoder (E): Extracts style features from a reference sketch. Alternatively, style can be sampled from a prior normal distribution to generate random variations.
- Class Recognizer (R) and Painter Identifier (I): These modules act as supervisors, ensuring the generated output maintains the correct semantic content and adheres to the target artist's style.
- Disentanglement Mechanism: The model separates shape from artistic "noise" using a combination of adversarial loss, class recognition loss, and an $L_1$ style reconstruction loss.
Style injection is achieved through Conditional Batch Normalization (CBN). A key innovation is the treatment of style as a continuous manifold, allowing for smooth interpolation between different artists' styles and the addition of noise to mimic natural human inconsistency.
Experimental Results
The model was evaluated using the SketchIME dataset, specifically focusing on 341 classes of geometric symbols. To test generalization, the training and test sets featured mutually exclusive "painters."
Visual Quality and Realism
The proposed model achieved a Fréchet Inception Distance (FID) of 12.175, significantly outperforming the state-of-the-art baseline, GANwriting (FID 23.548). This indicates a substantial improvement in the realism of the generated sketches.
Content and Style Accuracy
In "Class Extension" tests, where the model generated symbols not seen during training but composed of known geometric primitives, it achieved a content recognition accuracy (C_ACC) of over 96%. Additionally, the model showed superior performance in painter recognition accuracy (P_ACC), demonstrating its ability to capture the unique "signature" of an artist's hand.
Conclusion
This work demonstrates that disentangling geometric style from content is a viable path for high-quality sketch generation. The model's ability to perform zero-shot component recombination allows for the expansion of datasets into entirely new categories. Future research will focus on utilizing multiple reference images to build more comprehensive "painter profiles" and integrating synthetic data into downstream recognition pipelines to mitigate data scarcity in niche domains.
Summary updated