Abstract
Transfer learning is a cornerstone of modern computer vision, yet its effectiveness is often hampered by the "domain gap" between generic source datasets and specialized target tasks. This paper introduces a quantized similarity index that measures the visual information shared between datasets by counting correlated "atoms" in a sparse dictionary. By leveraging this index to select relevant intermediate datasets for a hierarchical transfer learning strategy, the authors demonstrate significant accuracy improvements in biometric tasks, such as hand-vein recognition, where traditional single-step transfer from generic models often fails.
Introduction
In contemporary deep learning, the standard practice is to take a Deep Convolutional Neural Network (DCNN) pre-trained on a massive, generic dataset like ImageNet (ILSVRC) and fine-tune it for a specific task. However, when the source domain (e.g., natural images of animals and vehicles) differs fundamentally from the target domain (e.g., infrared images of hand veins), the learned features often fail to generalize.
This paper addresses the lack of a quantitative metric to measure pattern similarity between datasets. The authors propose a mathematical framework to identify the "visual vocabulary" shared between domains, allowing researchers to predict transfer learning success and implement a more effective "hierarchical" transfer strategy.
Methods
The authors' methodology centers on two primary innovations:
- Quantized Image Pattern Similarity Index: Using sparse representation and dictionary learning, images are decomposed into fundamental "atoms." By constructing a structural sparse dictionary that distinguishes shared patterns from class-specific ones, the authors can calculate a "Shared Ratio." A higher ratio indicates that two datasets share more fundamental visual building blocks.
- Hierarchical Transfer Learning: Instead of a direct transfer from a generic model to a niche task, the authors introduce an intermediate stage. They use the similarity index to identify a "Data I" dataset that is more similar to the target "Data II." The DCNN is first re-trained on Data I to create a specialized feature extractor before the final application to the target task.
Datasets and Experimental Setup
The study validates the similarity index and hierarchical approach using a variety of datasets:
- Generic Source: ILSVRC (ImageNet).
- Intermediate Biometric Data: VGG Face, LFW (Labeled Faces in the Wild), PolyU NIR face, PolyU palmprint, and a fingerprint database.
- Target Task: A custom hand-dorsa vein database containing 2,000 images from 200 subjects, captured at an 850nm wavelength.
Key Results
The experimental results confirm that the similarity index is a robust predictor of transfer learning performance. The index correctly identified that fingerprints and hand veins share high similarity (56% shared atoms), whereas veins and the generic ILSVRC dataset share only 4%.
As shown in the performance data, the hierarchical approach yields dramatic improvements. In gender classification using hand veins, a VGG model transferred directly from ILSVRC achieved only 26.8% accuracy. By first re-training on the more similar PolyU palmprint dataset, accuracy rose to 93.5%. Similar trends were observed across identification and face recognition tasks.
Conclusion
This research demonstrates that "quality over quantity" is a vital principle in transfer learning. A smaller, highly similar dataset can serve as a far more effective stepping stone than a massive but dissimilar one. The proposed similarity index provides a rigorous mathematical framework for selecting these intermediate datasets, turning intuitive notions of visual similarity into a quantifiable metric for optimizing deep learning pipelines. Future work may focus on automating the search for optimal intermediate datasets within large-scale libraries.
Summary updated