[Paper Review] GCD-DDPM: A Generative Change Detection Model Based on Difference-Feature Guided DDPM

Yihan Wen, Xianping Ma+2 more|arXiv (Cornell University)|Jun 6, 2023

Remote-Sensing Image Classification4 citations

TL;DR

This paper proposes GCD-DDPM, a novel generative change detection model that leverages a Denoising Diffusion Probabilistic Model (DDPM) with difference-feature guided training to directly generate high-accuracy change detection maps. By integrating multi-scale difference conditional encoding and a noise suppression-based semantic enhancer, GCD-DDPM improves edge localization and feature discrimination, achieving state-of-the-art performance on four remote sensing datasets with significant gains in F1-score and IoU.

ABSTRACT

Deep learning (DL)-based methods have recently shown great promise in bitemporal change detection (CD). Existing discriminative methods based on Convolutional Neural Networks (CNNs) and Transformers rely on discriminative representation learning for change recognition while struggling with exploring local and long-range contextual dependencies. As a result, it is still challenging to obtain fine-grained and robust CD maps in diverse ground scenes. To cope with this challenge, this work proposes a generative change detection model called GCD-DDPM to directly generate CD maps by exploiting the Denoising Diffusion Probabilistic Model (DDPM), instead of classifying each pixel into changed or unchanged categories. Furthermore, the Difference Conditional Encoder (DCE), is designed to guide the generation of CD maps by exploiting multi-level difference features. Leveraging the variational inference (VI) procedure, GCD-DDPM can adaptively re-calibrate the CD results through an iterative inference process, while accurately distinguishing subtle and irregular changes in diverse scenes. Finally, a Noise Suppression-based Semantic Enhancer (NSSE) is specifically designed to mitigate noise in the current step's change-aware feature representations from the CD Encoder. This refinement, serving as an attention map, can guide subsequent iterations while enhancing CD accuracy. Extensive experiments on four high-resolution CD datasets confirm the superior performance of the proposed GCD-DDPM. The code for this work will be available at https://github.com/udrs/GCD.

Motivation & Objective

• To address the limitations of existing deep learning models in capturing long-range dependencies and preserving fine-grained local details in change detection.
• To overcome the shortcomings of pre-trained DDPMs that rely on single forward passes and miss iterative refinement benefits.
• To develop an end-to-end generative framework that directly produces high-quality change detection maps using variational inference and guided feature learning.
• To enhance feature representation and edge accuracy through multi-scale difference feature encoding and noise suppression.

Proposed method

• The model employs a denoising diffusion probabilistic model trained via variational inference to generate change detection maps iteratively, improving feature representation over time.
• A novel difference-feature guided conditional encoder integrates multi-level feature differences from pre- and post-change images to enhance change-aware representation.
• An adaptive calibration mechanism dynamically fuses difference features across three scales (SCALE1–SCALE3) to improve detection of small and irregular objects.
• A noise suppression-based semantic enhancer (NSSE) leverages prior knowledge from the current step to suppress high-frequency noise in the generated change maps.
• The model uses attention mechanisms to fuse information before applying pixel-wise difference operations, enabling dynamic localization of change regions.
• End-to-end training eliminates the need for pre-training, allowing direct optimization of the generative process for change detection.

Experimental results

Research questions

• RQ1 Can a diffusion-based generative model outperform conventional CNN and Transformer-based models in capturing both global context and local edge details in change detection?
• RQ2 How does multi-scale difference feature encoding affect the accuracy and robustness of change detection maps?
• RQ3 To what extent does the proposed noise suppression-based semantic enhancer improve the quality of generated change maps?
• RQ4 Does end-to-end training of a DDPM for change detection yield better performance than using pre-trained DDPMs as feature extractors?

Key findings

• On the WHU dataset, GCD-DDPM achieved an F1-score of 92.54% and IoU of 86.52%, outperforming the ablated model without NSSE by 1.02% and 1.83% respectively.
• The ablation study confirmed that incorporating both medium-scale (SCALE2) and small-scale (SCALE3) difference features significantly improved performance over using only SCALE3.
• The complete GCD-DDPM model achieved an overall accuracy of 99.39% on the WHU dataset, demonstrating strong generalization across diverse scenes.
• Compared to DDPM-CD, GCD-DDPM reduced computational complexity from 838.6 GFLOPs to 269.5 GFLOPs, offering a more efficient solution despite higher parameter count.
• Extensive experiments on four datasets (CDD, WHU, Levier, GVLM) confirmed that GCD-DDPM consistently outperforms state-of-the-art models in F1-score and IoU.
• Visual results showed that GCD-DDPM generated more discriminative and change-aware feature maps at all levels, preserving fine structures and reducing localization errors.