Semi-U-Net: A Lightweight Deep Neural Network for Subject-Sensitive Hashing of HRRS Images

Kaimeng Ding, Shoubao Su, Nan Xu, Tingting JiangIEEE Access 2021 · DOI

Introduction

High-resolution remote sensing (HRRS) images are indispensable for modern urban planning, environmental monitoring, and disaster response. However, ensuring the integrity of these images is complex. Traditional cryptographic hashes are overly sensitive to bit-level changes, such as those caused by standard image compression, while digital watermarking often alters the original data.

Subject-sensitive hashing provides a solution by generating hashes based on semantic content—the "subjects" of interest, such as buildings—rather than raw pixel values. This allows the system to ignore harmless modifications like brightness adjustments while detecting malicious tampering. Despite its potential, existing deep learning models for this task are often too computationally intensive for deployment on edge devices like drones.

The Semi-U-Net Architecture

To address these efficiency bottlenecks, researchers proposed Semi-U-Net, a lightweight deep neural network optimized for HRRS hashing. The architecture introduces three key innovations:

  1. Asymmetric Semi-U-Shaped Structure: While standard U-Nets use symmetrical encoders and decoders for segmentation, Semi-U-Net employs a 3-level encoder and a minimal 1-level decoder. This reduction prevents the model from extracting redundant features, which enhances robustness against noise and significantly lowers complexity.
  2. Depthwise Separable Convolutions: By replacing standard convolutions with depthwise separable ones, the model splits computation into filtering and combining stages. This reduces computational costs by nearly 90% with negligible impact on accuracy.
  3. Multi-Scale Data Input: The network samples the original image at multiple resolutions, feeding them directly into different encoder stages to maintain global context without increasing model depth.

Performance Benchmarks

The Semi-U-Net was evaluated against state-of-the-art models like MUM-net and MultiResUnet using datasets derived from the WHU Building Dataset and GaoFen-2 (GF-2) imagery.

Metric Semi-U-Net MUM-net MultiResUnet
Model Size 5.38 MB 146.64 MB 81.2 MB
Speed (FPS) 88.6 42.1 30.6
FLOPs 0.91 M 25.54 M 14.5 M

The results demonstrate that Semi-U-Net is approximately 1/27th the size of MUM-net and operates 2.1x faster, making it highly suitable for real-time applications.

Robustness and Tampering Detection

A critical requirement for subject-sensitive hashing is the ability to distinguish between "content-preserving" operations and "malicious tampering."

Examples of original (a) and tampered (b) HRRS images used to test subject-sensitive hashing sensitivity.

In testing, Semi-U-Net successfully detected semantic changes—such as the addition or deletion of buildings—while remaining robust to PNG compression and digital watermarking. The researchers found that the model's reduced complexity actually aided stability, as it focused strictly on the primary subjects rather than fluctuating with minor background noise.

Conclusion

Semi-U-Net represents a significant step toward practical, real-time integrity authentication for remote sensing. By prioritizing architectural efficiency and semantic focus, it achieves high performance on a fraction of the hardware requirements of previous models. Future work will explore deployment on mobile ARM processors and further optimization of the hash comparison process.

Summary updated

Abstract

As a special case of perceptual hashing algorithm, subject-sensitive hashing can realize “subject-biased” integrity authentication of high resolution remote sensing (HRRS) images, which overcomes the deficiencies of existing integrity authentication technologies. However, the existing deep neural network for subject-sensitive hashing have disadvantages such as high model complexity and low computational efficiency. In this paper, we propose an efficient and lightweight deep neural network named Semi-U-net to achieve efficient subject-sensitive hashing. The proposed Semi-U-net realizes the lightweight of the network from three aspects: First, considering the general process of perceptual hashing, it adopts a semi-u-shaped structure, which simplify the model structure and prevent the model from extracting too much redundant information to enhance the robustness of the algorithm; Second, the number of model parameters and the computational cost are significantly reduced by using deep separable convolution in the entire asymmetric network; Third, the number of model parameters is further compressed by using the dropout layer several times. The experimental results show that the size of our Semi-U-Net model is only 5.38M, which is only 1/27 of MUM-net and 1/15 of MultiResUnet. The speed of the Semi-U-Net based subject-sensitive hashing algorithm is 88.6 FPS, which is 2.89 times faster than MultiResUnet based algorithm and 2.1 times faster than MUM-net Based Algorithm. FLOPs of Semi-U-net is only 1/28 of MUM-net and 1/16 of MultiResUnet.
BibTeX
@inproceedings{ding2021semi,
  author = {Kaimeng Ding and Shoubao Su and Nan Xu and Tingting Jiang},
  title = {{Semi-U-Net: A Lightweight Deep Neural Network for Subject-Sensitive Hashing of HRRS Images}},
  booktitle = {IEEE Access},
  year = {2021},
  doi = {10.1109/access.2021.3074055},
}
Figure