Abstract
Extracting informative image features and learning effective approximate hashing functions are two crucial steps in image retrieval. Conventional methods often study these two steps separately, e.g., learning hash functions from a predefined hand-crafted feature space. Meanwhile, the bit lengths of output hashing codes are preset in most previous methods, neglecting the significance level of different bits and restricting their practical flexibility. To address these issues, we propose a supervised learning framework to generate compact and bit-scalable hashing codes directly from raw images. We pose hashing learning as a problem of regularized similarity learning. Specifically, we organize the training images into a batch of triplet samples, each sample containing two images with the same label and one with a different label. With these triplet samples, we maximize the margin between matched pairs and mismatched pairs in the Hamming space. In addition, a regularization term is introduced to enforce the adjacency consistency, i.e., images of similar appearances should have similar codes. The deep convolutional neural network is utilized to train the model in an end-to-end fashion, where discriminative image features and hash functions are simultaneously optimized. Furthermore, each bit of our hashing codes is unequally weighted so that we can manipulate the code lengths by truncating the insignificant bits. Our framework outperforms state-of-the-arts on public benchmarks of similar image search and also achieves promising results in the application of person re-identification in surveillance. It is also shown that the generated bit-scalable hashing codes well preserve the discriminative powers with shorter code lengths.
Introduction
In large-scale image retrieval, efficiency is achieved by converting images into compact binary "hash codes." However, traditional systems face two primary bottlenecks: a "Pipeline Gap" where hand-crafted features are not optimized for hashing, and "Rigid Code Lengths" that require separate models for different bit requirements. This paper introduces a supervised learning framework to generate compact and bit-scalable hashing codes directly from raw images, optimizing feature extraction and hashing simultaneously.
Methodology: Deep Regularized Similarity Comparison Hashing (DRSCH)
The proposed DRSCH framework utilizes a deep convolutional neural network (CNN) to train the model in an end-to-end fashion. The architecture, illustrated in Figure 4, bridges the gap between feature learning and hash function optimization.
Key Technical Innovations
- Triplet-Based Similarity Learning: The model is trained on triplets (query, similar, dissimilar) to maximize the margin between matched and mismatched pairs in Hamming space.
- Adjacency Regularization: A Laplacian regularization term ensures that images with similar visual appearances maintain their neighborhood relationships, preserving local topology and preventing over-focusing on class labels.
- Bit-Scalable Layer: An element-wise weighting layer learns to assign different importance levels to each bit. This allows for "bit-scalability," where truncating a long code (e.g., 64 bits) to a shorter one (e.g., 16 bits) still preserves significant discriminative power.
- Differentiable Approximation: To enable backpropagation through discrete binary values, a "tanh-like" activation function is used to approximate the sign function during training.
Experimental Results
The model was evaluated on standard benchmarks including MNIST, CIFAR-10/20, NUS-WIDE, and the CUHK03 person re-identification dataset.
As shown in Figure 6, DRSCH consistently outperforms state-of-the-art methods like DSRH and KSH-CNN across all bit lengths on the MNIST dataset, achieving a Mean Average Precision (MAP) of 98.09% at 64 bits. On the NUS-WIDE dataset, the bit-scalable version demonstrated that using only the top 8 bits of a 64-bit model could achieve competitive results (58.77% MAP), proving the effectiveness of the bit-weighting strategy. Furthermore, the model achieved promising results in person re-identification, outperforming specialized algorithms on the CUHK03 dataset.
Conclusion
The DRSCH framework successfully integrates feature learning and hashing into a single, end-to-end trainable model. By introducing bit-scalability and adjacency regularization, the authors provide a flexible and robust solution for large-scale image retrieval and surveillance applications. Future work may involve leveraging specific image attributes and incorporating feedback loops to further enhance retrieval accuracy.
Summary updated