Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Weakly Supervised Change Detection via Knowledge Distillation and Multiscale Sigmoid Inference

Binghao Lu, Caiwen Ding, Jinbo Bi, Dongjin Song

TL;DR

A novel weakly supervised change detection technique via Knowledge Distillation and Multiscale Sigmoid Inference (KD-MSI) that leverages image-level labels that significantly outperforms the state-of-the-art.

Abstract

Change detection, which aims to detect spatial changes from a pair of multi-temporal images due to natural or man-made causes, has been widely applied in remote sensing, disaster management, urban management, etc. Most existing change detection approaches, however, are fully supervised and require labor-intensive pixel-level labels. To address this, we develop a novel weakly supervised change detection technique via Knowledge Distillation and Multiscale Sigmoid Inference (KD-MSI) that leverages image-level labels. In our approach, the Class Activation Maps (CAM) are utilized not only to derive a change probability map but also to serve as a foundation for the knowledge distillation process. This is done through a joint training strategy of the teacher and student networks, enabling the student network to highlight potential change areas more accurately than teacher network based on image-level labels. Moreover, we designed a Multiscale Sigmoid Inference (MSI) module as a post processing step to further refine the change probability map from the trained student network. Empirical results on three public datasets, i.e., WHU-CD, DSIFN-CD, and LEVIR-CD, demonstrate that our proposed technique, with its integrated training strategy, significantly outperforms the state-of-the-art.

Weakly Supervised Change Detection via Knowledge Distillation and Multiscale Sigmoid Inference

TL;DR

A novel weakly supervised change detection technique via Knowledge Distillation and Multiscale Sigmoid Inference (KD-MSI) that leverages image-level labels that significantly outperforms the state-of-the-art.

Abstract

Change detection, which aims to detect spatial changes from a pair of multi-temporal images due to natural or man-made causes, has been widely applied in remote sensing, disaster management, urban management, etc. Most existing change detection approaches, however, are fully supervised and require labor-intensive pixel-level labels. To address this, we develop a novel weakly supervised change detection technique via Knowledge Distillation and Multiscale Sigmoid Inference (KD-MSI) that leverages image-level labels. In our approach, the Class Activation Maps (CAM) are utilized not only to derive a change probability map but also to serve as a foundation for the knowledge distillation process. This is done through a joint training strategy of the teacher and student networks, enabling the student network to highlight potential change areas more accurately than teacher network based on image-level labels. Moreover, we designed a Multiscale Sigmoid Inference (MSI) module as a post processing step to further refine the change probability map from the trained student network. Empirical results on three public datasets, i.e., WHU-CD, DSIFN-CD, and LEVIR-CD, demonstrate that our proposed technique, with its integrated training strategy, significantly outperforms the state-of-the-art.
Paper Structure (21 sections, 6 equations, 5 figures, 4 tables, 1 algorithm)

This paper contains 21 sections, 6 equations, 5 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Fully-supervised change detection
  4. Weakly-supervised semantic segmentation
  5. Weakly supervised change detection
  6. Knowledge distillation
  7. Proposed Method
  8. The Knowledge Distillation Framework
  9. Siamese Teacher Network
  10. Siamese Student Network
  11. Learning Objective
  12. Multiscale Sigmoid Inference
  13. Multiscale Inference
  14. Multiscale Sigmoid Inference
  15. Experiments
  16. ...and 6 more sections

Figures (5)

  • Figure 1: Comparison of change detection methods with pixel-level supervision (third column) vs. image-level supervision (fourth column).
  • Figure 2: Weakly supervised change detection network with a knowledge distillation framework. Both student and teacher networks are Siamese networks. $\bigoplus$ stands for feature combination methods to combine two time period feature maps, e.g., subtraction, absolute subtraction, and concatenation. Both teacher and student networks are trained concurrently. G stands for one channel feature map, GAP stands for global average pooling, CAM stands for class activation map, $\sigma$ is sigmoid activation function
  • Figure 3: LEVIR-CD test dataset performance comparison.
  • Figure 4: WHU-CD test dataset performance comparison.
  • Figure 5: Visualization of change probability map on WHU-CD dataset, where column(a) denote pre-event images, column(b) represents post-event images, column(c) represents ground truth, column(d) represents CAM from teacher model, column(e) represents change probability map from student model, column(f) stands for change probability from student model with MSI.