An Application of Ternary Hash Retrieval Method for Remote Sensing Images in Panoramic Video

Weipeng Jing, Duanpeng Zhang, Houbing Song

Abstract

In the context of increasing reliance on drones and autonomous vehicles for real-time panoramic mapping, efficient image retrieval from massive remote sensing databases has become a critical challenge. This paper presents Deep Feature Based Ternary Hashing (DFTH), a novel one-stage end-to-end deep hashing framework. Unlike traditional two-stage methods or those relying on pairwise labels, DFTH utilizes ternary labels (triplets) to capture complex semantic relationships and optimizes image features and hash codes simultaneously. By employing a continuous incremental transform to handle discrete optimization, the proposed method achieves a 3-4% improvement in Mean Average Precision (MAP) over current state-of-the-art techniques on standard benchmarks. These results demonstrate the potential of DFTH for high-speed, high-accuracy retrieval in real-time remote sensing applications.

Introduction

The proliferation of drones and autonomous systems has led to a surge in demand for high-quality, real-time panoramic maps derived from vast remote sensing image databases. Traditional retrieval methods, such as text-based indexing or manual feature extraction, struggle with the high dimensionality and scale of modern data. Hashing techniques, which compress high-dimensional data into short binary codes, offer a scalable solution for rapid similarity search in Hamming space. However, many existing deep hashing methods are limited by two-stage learning processes or the use of simple pairwise labels that fail to capture nuanced semantic similarities.

Proposed Method: Deep Feature Based Ternary Hashing (DFTH)

The authors propose DFTH to address these limitations through a unified, end-to-end learning framework. The methodology centers on three key innovations:

End-to-End Feature and Hash Learning

DFTH utilizes a single CNN-F architecture to learn visual features and generate hash codes simultaneously. This one-stage approach ensures that the learned features are specifically tailored for binary representation, avoiding the information loss associated with decoupled feature extraction and hashing stages.

Ternary Label Likelihood

The model moves beyond binary pairwise comparisons by adopting a triplet-based approach. Each training instance consists of a query image, a positive (similar) image, and a negative (dissimilar) image. The loss function is designed to maximize the likelihood that the query image is significantly closer to the positive image than to the negative image in the resulting binary space.

The DFTH architecture showing the triplet-based learning process with shared weights across query, positive, and negative branches.

Continuous Incremental Transform

To overcome the mathematical difficulty of training a network to output discrete binary bits (0s and 1s), the authors implement a "continuation method." They employ a scaled tanh function that begins as a smooth continuous function and gradually sharpens into a step function as training progresses. This allows the network to optimize over continuous values before being constrained to discrete hash codes.

Experimental Results

The DFTH method was evaluated on the CIFAR-10 and NUS-WIDE datasets to benchmark its performance against established hashing algorithms.

  • Accuracy Improvements: On CIFAR-10, DFTH showed a 3% to 4% improvement in Mean Average Precision (MAP) for 24-bit and 32-bit hash codes compared to state-of-the-art methods like DPSH.
  • Consistency: The method achieved an average MAP of 0.753, significantly outperforming traditional CNN-based hashing methods which typically range between 0.53 and 0.60.
  • Training Efficiency: The use of a positive margin ($\alpha$) in the loss function helped prevent vanishing gradients, ensuring faster convergence and distinct separation between semantic classes.

Conclusion

The Deep Feature Based Ternary Hashing (DFTH) method offers a robust and efficient solution for remote sensing image retrieval. By integrating feature learning and hash coding into a single stage and utilizing ternary labels, the framework achieves superior retrieval accuracy. Future work will focus on adopting more complex backbones, such as ResNet or Transformers, and further optimizing retrieval speeds for real-time drone navigation and panoramic video applications.

Summary updated

Abstract

With the development of multimedia technology, the application of panoramic video is emerging, and there are specific application scenarios in the field of drones. Remote sensing images have tremendous applications in data acquisition and processing in the panoramic maps used to construct the panoramic videos. Efficient remote sensing image retrieval can greatly improve the efficiency of image utilization and get a fast response in the construction of panoramic video. Thus, how to achieve fast and efficient remote sensing image retrieval becomes increasingly important. Scholars have explored a number of ways to address this problem, from traditional manual feature-based retrieval methods to deep hash retrieval methods that today can cope with large data volumes. The hash method is a way to encode high-dimensional data into a set of binary codes while maintaining similarity between images. In this paper, we propose an end-to-end ternary label depth hashing method that is capable of both feature learning and hash coding processes in one-stage. The experimental results show that the algorithm has high accuracy and good retrieval effect.
BibTeX
@article{jing2020application,
  author = {weipeng jing and Weipeng Jing and Duanpeng Zhang and Houbing Song},
  title = {{An Application of Ternary Hash Retrieval Method for Remote Sensing Images in Panoramic Video}},
  journal = {IEEE Access},
  year = {2020},
  doi = {10.1109/access.2020.3006103},
}
Figure