LogoNet: a fine-grained network for instance-level logo sketch retrieval
This work addresses a specific challenge in sketch-based image retrieval for logo identification, but it is incremental as it builds on existing sketch retrieval methods with a new dataset and architecture.
The authors tackled the problem of instance-level logo sketch retrieval by constructing the first publicly available dataset with 2k logo instances and over 9k sketches, and developed LogoNet, a fine-grained CNN architecture with hybrid attention, which demonstrated effectiveness in experiments.
Sketch-based image retrieval, which aims to use sketches as queries to retrieve images containing the same query instance, receives increasing attention in recent years. Although dramatic progress has been made in sketch retrieval, few efforts are devoted to logo sketch retrieval which is still hindered by the following challenges: Firstly, logo sketch retrieval is more difficult than typical sketch retrieval problem, since a logo sketch usually contains much less visual contents with only irregular strokes and lines. Secondly, instance-specific sketches demonstrate dramatic appearance variances, making them less identifiable when querying the same logo instance. Thirdly, there exist several sketch retrieval benchmarking datasets nowadays, whereas an instance-level logo sketch dataset is still publicly unavailable. To address the above-mentioned limitations, we make twofold contributions in this study for instance-level logo sketch retrieval. To begin with, we construct an instance-level logo sketch dataset containing 2k logo instances and exceeding 9k sketches. To our knowledge, this is the first publicly available instance-level logo sketch dataset. Next, we develop a fine-grained triple-branch CNN architecture based on hybrid attention mechanism termed LogoNet for accurate logo sketch retrieval. More specifically, we embed the hybrid attention mechanism into the triple-branch architecture for capturing the key query-specific information from the limited visual cues in the logo sketches. Experimental evaluations both on our assembled dataset and public benchmark datasets demonstrate the effectiveness of our proposed network.