Abstract
To extract image features that fully express semantic information, reduce projection errors in hashing retrieval, and generate more compact binary hash codes, this research proposes a method based on DenseNet and improved Supervised Kernel Hashing (KSH). High-level semantic features are extracted using an optimized DenseNet. Kernel Principal Component Analysis (KPCA) is then applied to these features to capture non-linear information and minimize projection error. Finally, KSH is used for supervised learning to map features into Hamming space. Experimental results on Paris6K and LUNA16 datasets show that the proposed method achieves high mean Average Precision (mAP), reaching 89.2% and 92.9% respectively at 64-bit code lengths, with lower computational complexity than CNNH. The method is effective for both color natural images and grayscale medical images.
Introduction
As digital image repositories expand exponentially, traditional high-dimensional feature matching suffers from the "curse of dimensionality," leading to prohibitive storage requirements and slow query times. Hashing techniques address this by mapping images into compact binary codes in Hamming space, enabling rapid similarity searches via bitwise operations.
However, two primary challenges persist in hashing-based retrieval:
- Feature Representation: Shallow or hand-crafted features often fail to capture the deep semantic nuances required for accurate retrieval.
- Projection Error: The transition from high-dimensional continuous features to short binary codes often incurs significant information loss, particularly when data distributions are non-linear.
This paper introduces a novel framework that combines deep learning with non-linear projection and supervised hashing to overcome these hurdles.
Methodology
The proposed retrieval pipeline consists of three integrated stages designed to generate high-quality binary "fingerprints" for images.
1. Deep Feature Extraction via DenseNet
The framework utilizes a DenseNet architecture, optimized with Batch Normalization (BN) and the RMSProp optimizer, to extract high-level semantic descriptors. DenseNet's unique architecture, characterized by dense connectivity between layers, promotes feature reuse and mitigates the vanishing gradient problem, resulting in more expressive image representations than standard convolutional neural networks (CNNs).
2. Non-linear Dimensionality Reduction (KPCA)
To bridge the gap between high-dimensional features and binary codes, the authors introduce a Kernel Principal Component Analysis (KPCA) step. By projecting features into a higher-dimensional space before reduction, KPCA captures the underlying non-linear structures of the data. This significantly reduces the projection error that typically plagues linear hashing methods.
3. Improved Supervised Kernel Hashing (KSH)
The final stage employs a supervised hashing algorithm that leverages image labels to learn a mapping into Hamming space. By incorporating semantic labels during training, the resulting binary codes ensure that images with similar meanings are mapped to nearby points in the binary space, maximizing retrieval accuracy.
Experimental Evaluation
The versatility of the DenseNet-KPCA-KSH framework was tested on two distinct benchmarks:
- Paris6K: A dataset of 6,412 high-variance color images of Parisian landmarks, representing natural scene retrieval.
- LUNA16 (Lung Nodule Analysis 16): A medical dataset comprising 888 lung CT scans. This benchmark tests the model's precision on grayscale, low-contrast images where subtle textural differences are paramount.
Results and Analysis
The proposed method was benchmarked against several prominent hashing algorithms, including DNNH, CNNH, KSH, BRE, ITQ, and SH.
Retrieval Performance
The framework consistently outperformed competitors across various bit lengths (12 to 128 bits). Peak performance was observed at 64 bits, with an mAP of 89.2% on Paris6K and 92.9% on LUNA16. Notably, the performance on medical data exceeded that of natural images, highlighting the model's robustness in specialized domains.
Computational Efficiency
Efficiency is a key advantage of this modular approach. On the Paris6K dataset, the training time for a 64-bit code was approximately 108.36 seconds, outperforming the end-to-end CNNH model (118.94 seconds). Lower testing times further validate its suitability for real-time large-scale retrieval applications.
Conclusion
By combining the feature-extraction power of DenseNet with the non-linear projection capabilities of KPCA, this research provides a highly effective solution for image hash retrieval. The modular design improves accuracy and reduces projection error while maintaining competitive computational speeds. Future work will explore integrating the hashing layer directly into the DenseNet architecture for a fully end-to-end training process, potentially further optimizing the learned features for Hamming space.
Summary updated