Leveraging Fine-Grained Information and Noise Decoupling for Remote Sensing Change Detection

TL;DR

This work targets remote-sensing change detection by addressing noise that obscures fine-grained information in bi-temporal image pairs. It introduces FINO, a framework combining context-dependent learning to recover fine-grained features, brightness-aware and shape-aware modules to decouple task-agnostic noise, and a regularization gate to separate task-specific noise. The approach yields state-of-the-art performance on LEVIR-CD, WHU-CD, DSIFN-CD, and CDD2018 Lebe datasets, with consistent improvements in F1 and IoU and improved boundary delineation in qualitative results. By preserving local detail while suppressing noise, FINO enhances detection robustness, especially for small or densely packed changes, and the authors plan to release code for broader adoption.

Abstract

Change detection aims to identify remote sense object changes by analyzing data between bitemporal image pairs. Due to the large temporal and spatial span of data collection in change detection image pairs, there are often a significant amount of task-specific and task-agnostic noise. Previous effort has focused excessively on denoising, with this goes a great deal of loss of fine-grained information. In this paper, we revisit the importance of fine-grained features in change detection and propose a series of operations for fine-grained information compensation and noise decoupling (FINO). First, the context is utilized to compensate for the fine-grained information in the feature space. Next, a shape-aware and a brightness-aware module are designed to improve the capacity for representation learning. The shape-aware module guides the backbone for more precise shape estimation, guiding the backbone network in extracting object shape features. The brightness-aware module learns a overall brightness estimation to improve the model's robustness to task-agnostic noise. Finally, a task-specific noise decoupling structure is designed as a way to improve the model's ability to separate noise interference from feature similarity. With these training schemes, our proposed method achieves new state-of-the-art (SOTA) results in multiple change detection benchmarks. The code will be made available.

Figures (6)

• Figure 1: Conceptual illustration of proposed approach. (a) Two types of change detection pseudo-change. (b) Multi-scale fused features (far left) visualised with features learned by FINO (middle), with FINO predictions on the far right. (c) The core process of FINO. Context-dependent learning (CDL) compensates for fine-grained features, brightness-aware and shape-aware (BSA) perception to decouple task-agnostic noise, and regularization gate (REGA) decouples task-specific noise.
• Figure 2: Framework of the proposed network. FINO consists of CDL, shape-aware module, and REGA in tandem. The attention-based CDL adequately compensates for fine-grained information. The shape-aware module decouples the task-agnostic noise and guides the model to learn the object shape representation. REGA decouples the task-specific noise.
• Figure 3: FINO Detail Structure. FINO consists of CDL, shape-aware module and REGA. First, CDL compensates the fine-grained information in $C_i$ through $Z_{i+1}$ to obtain the compensated feature $T_i$. Then, the shape-aware module learns the shape to obtain $M_i$ and enriched feature $H_i$. In REGA, $H_i$ and the bitemporal features $X_{t_1,i} X_{t_2,i}$ pass through a gated structure to obtain the change feature $Z_i$ of the current layer, where $M_i$ is used as a regularization term for the gated structure to improve the robustness to pseudo changes.
• Figure 4: Qualitative results of different CD methods on LEVIR-CD. (a) $T_1$ image. (b) $T_2$ image. (c) FC-EF. (d) FC-Siam-Diff. (e) FC-Siam-Conc. (f) DTCDSCN. (g) BIT. (h) ChangeFormer. (i) FINO(ours). (j) Ground truth. where the first row shows the comparison of detection results for dense objects and the second row shows the comparison of detection results for multi-scale objects. The yellow box is the detection effect of small objects, the red box is the detection effect of large objects, and the blue box is the detection effect of object edges.
• Figure 5: Qualitative results of different CD methods of targets at different scales. (a) $T_1$ image. (b) $T_2$ image. (c) FC-Siam-diff. (d) BiT. (e) ChangeFormer (f) USSFC-Net. (g) SARAS-Net. (h) Changer. (i) FINO(ours). (j) Ground truth.