Related Experiment Video
Updated: Jun 27, 2025
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
524
Zero-Shot Sketch-Based Image Retrieval Using StyleGen and Stacked Siamese Neural Networks
Venkata Rama Muni Kumar Gopu1, Madhavi Dunna1
1Department of Electrical, Electronics and Communication Engineering (EECE), Gitam School of Technology, Gitam Deemed to be University, Rushikonda, Visakhapatnam 530045, India.
Journal of Imaging
|April 26, 2024
Summary
This study introduces StyleGen, a novel method for sketch-based image retrieval (SBIR). StyleGen generates realistic images from sketches, significantly improving retrieval accuracy in zero-shot scenarios by bridging the domain gap.
Area of Science:
- Computer Vision
- Machine Learning
- Artificial Intelligence
Background:
- Sketch-based image retrieval (SBIR) is challenging due to the domain gap between ambiguous sketches and natural images.
- Zero-shot SBIR requires retrieving images from classes not seen during model training, exacerbating the domain gap.
Purpose of the Study:
- To propose StyleGen, an elegant retrieval methodology for generating candidate images that match the domain of repository images.
- To reduce the domain gap in sketch-based image retrieval tasks.
Main Methods:
- Utilizes a two-stage neural network architecture: the stacked Siamese network.
- Employs StyleGen to generate fake candidate images aligned with the target image domain.
Main Results:
- Demonstrates marked performance improvement on TU-Berlin Extended and Sketchy Extended datasets.
- Achieves superior results compared to current state-of-the-art approaches in SBIR.
- Maintains generalizability of the retrieval approach.
Conclusions:
- StyleGen effectively bridges the domain gap in SBIR.
- The stacked Siamese network architecture provides outstanding retrieval performance.
- The proposed methodology significantly advances zero-shot SBIR capabilities.
