Abstract
Due to the impressive learning power, deep learning has achieved a remarkable performance in supervised hash function learning. In this paper, we propose a novel asymmetric supervised deep hashing method to preserve the semantic structure among different categories and generate the binary codes simultaneously. Specifically, two asymmetric deep networks are constructed to reveal the similarity between each pair of images according to their semantic labels. The deep hash functions are then learned through two networks by minimizing the gap between the learned features and discrete codes. Furthermore, since the binary codes in the Hamming space also should keep the semantic affinity existing in the original space, another asymmetric pairwise loss is introduced to capture the similarity between the binary codes and real-value features. This asymmetric loss not only improves the retrieval performance, but also contributes to a quick convergence at the training phase. By taking advantage of the two-stream deep structures and two types of asymmetric pairwise functions, an alternating algorithm is designed to optimize the deep features and high-quality binary codes efficiently. Experimental results on three real-world datasets substantiate the effectiveness and superiority of our approach as compared with state-of-the-art.
Introduction
Deep hashing has emerged as a powerful technique for efficient image retrieval by mapping high-dimensional data into compact binary codes. While supervised deep learning has dominated this field, traditional symmetric architectures—where two images pass through identical networks to compute Hamming distances—often encounter optimization bottlenecks. The discrete constraints of binary codes make the resulting objective functions highly non-convex and difficult to optimize directly, frequently leading to performance degradation during relaxation.
This paper introduces Dual Asymmetric Deep Hashing Learning (DADH), a framework designed to overcome these limitations. By employing two distinct neural network streams and a novel asymmetric pairwise loss, DADH bridges the gap between continuous real-valued features and discrete binary codes, ensuring both high retrieval accuracy and rapid convergence.
Methodology
The DADH framework is built upon three core components: dual asymmetric networks, a consistent binary code target, and an asymmetric pairwise loss function.
Dual Asymmetric Networks
Unlike symmetric Siamese-style networks, DADH utilizes two separate Convolutional Neural Network (CNN) streams (based on the CNN-F architecture) to learn distinct hash functions. This asymmetry facilitates the optimization of discrete binary constraints, which is typically a mathematical bottleneck in symmetric models.
Consistent Binary Codes and Asymmetric Loss
A key distinction from previous asymmetric methods is that DADH forces both networks to map to a single, consistent set of binary codes for the training data. The central innovation is the Asymmetric Pairwise Loss, which computes the inner product between the real-valued network outputs (approximated via a tanh function) and the discrete binary codes. This approach acts as a robust regularizer, preserving semantic similarity while minimizing quantization errors typical of purely discrete comparisons.
Alternating Optimization
To handle the NP-hard nature of discrete optimization, the authors employ an alternating algorithm. The process iteratively freezes the binary codes to update the networks via backpropagation, and then freezes the network weights to update the binary codes bit-by-bit using a closed-form discrete solution.
Experimental Evaluation
The performance of DADH was evaluated on three benchmark datasets: IAPR TC-12 (20,000 images), MIRFLICKR-25K (25,000 images), and CIFAR-10 (60,000 images).
Retrieval Performance
DADH consistently outperformed both traditional hashing methods (e.g., LSH, ITQ) and state-of-the-art deep hashing baselines such as DPSH, ADSH, and DAPH across various code lengths. For instance, on CIFAR-10 with 48-bit codes, DADH achieved a Mean Average Precision (MAP) of 83.90%, significantly higher than DAPH's 70.28%. On IAPR TC-12, DADH reached 55.39% MAP at 48 bits, a nearly 10% improvement over previous asymmetric approaches.
Convergence and Efficiency
The asymmetric loss formulation contributes to exceptionally fast convergence. Empirical results show that the alternating optimization algorithm stabilizes in fewer than 30 iterations. Ablation studies confirmed that the continuous-to-discrete loss terms are critical for both speed and reaching superior local optima.
Conclusion
Dual Asymmetric Deep Hashing Learning demonstrates that asymmetry in network architecture and loss functions can significantly enhance the learning of discrete hash codes. By bridging the continuous-discrete divide, DADH achieves state-of-the-art results with high training efficiency. Future work may explore reducing the computational overhead of bit-by-bit updates for massive datasets and extending the framework to unsupervised scenarios where label information is unavailable.
Summary updated