摘要
高精度特征提取模型对于变化检测（Change Detection, CD）至关重要。过去，许多基于深度学习的监督CD方法通过大量标注的双时相图像学习识别变化特征模式，然而，标注双时相遥感图像非常昂贵且耗时；因此，我们提出了一种基于一致性正则化的粗到精半监督CD方法（Coarse-to-Fine Semi-Supervised Change Detection, C2F-SemiCD），该方法包括一个具有多尺度注意力机制的粗到精CD网络（Coarse-to-Fine Network, C2FNet）和一种半监督更新方法。其中，C2FNet网络通过多尺度特征融合、通道注意力机制、空间注意力机制、全局上下文模块、特征细化模块、初始聚合模块和最终聚合模块，逐步从粗粒度到细粒度完成变化特征的提取。半监督更新方法采用了均值教师方法，学生模型的参数通过指数移动平均（Exponential Moving Average, EMA）方法更新到教师模型的参数。通过对三个数据集进行广泛的实验以及细致的消融研究，包括跨数据集交叉实验，我们验证了所提出的C2F-SemiCD方法在有效性和效率方面的显著优势。代码将在以下地址公开：https://github.com/ChengxiHAN/C2F-SemiCDand-C2FNet。
代码仓库
chengxihan/c2f-semicd-and-c2f-cdnet
官方
pytorch
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| change-detection-on-cdd-dataset-season-1 | C2FNet | F1: 95.93 F1-Score: 95.93 IoU: 92.18 KC: 95.39 Overall Accuracy: 99.04 Precision: 95.46 Recall: 96.41 |
| change-detection-on-dsifn-cd | C2FNet | F1: 64.03 IoU: 47.09 KC: 55.62 Overall Accuracy: 86.19 Precision: 57.45 Recall: 72.31 |
| change-detection-on-googlegz-cd | C2FNet | F1: 86.86 IoU: 76.77 KC: 82.48 Overal Accuracy: 93.43 Precision: 85.46 Recall: 88.31 |
| change-detection-on-levir | C2FNet | F1: 79.15 IoU: 65.50 KC: 78.25 OA: 98.26 Prcision: 77.19 Recall: 81.22 |
| change-detection-on-levir-cd | C2FNet | F1: 91.83 F1-score: 99.18 IoU: 93.69 Overall Accuracy: 90.04 Precision: 84.89 Recall: 91.40 |
| change-detection-on-s2looking | C2FNet | F1: 62.83 F1-Score: 62.83 IoU: 45.80 KC: 62.44 OA: 99.22 Precision: 74.84 Recall: 54.14 |
| change-detection-on-sysu-cd | C2FNet | F1: 77.97 IoU: 63.89 KC: 70.87 OA: 89.25 Precision: 75.44 Recall: 80.67 |
| change-detection-on-whu-cd | C2FNet | F1: 94.36 IoU: 89.33 KC: 94.14 Overall Accuracy: 99.56 Precision: 96.57 Recall: 92.26 |
| semi-supervised-change-detection-on-levir-cd | C2F-SemiCD | F1: 89.97 IoU: 81.76 KC: 89.44 OA: 98.99 Precision: 91.45 Recall: 88.53 |
| semi-supervised-change-detection-on-levir-cd-1 | C2F-SemiCD | F1: 90.80 IoU: 83.15 KC: 90.31 OA: 99.08 Precision: 92.44 Recall: 89.22 |
| semi-supervised-change-detection-on-levir-cd-2 | C2F-SemiCD | F1: 91.16 IoU: 83.75 KC: 90.69 OA: 99.12 Precision: 93.26 Recall: 89.15 |
| semi-supervised-change-detection-on-levir-cd-3 | C2F-SemiCD | F1: 91.67 IoU: 84.62 KC: 91.23 OA: 99.17 Precision: 93.41 Recall: 89.99 |
| semi-supervised-change-detection-on-whu-10 | C2F-SemiCD | F1: 86.58 IoU: 76.33 KC: 86.03 OA: 98.94 Precision: 87.35 Recall: 85.81 |
| semi-supervised-change-detection-on-whu-20 | C2F-SemiCD | F1: 90.07 IoU: 81.93 KC: 89.66 OA: 99.23 Precision: 91.83 Recall: 88.36 |
| semi-supervised-change-detection-on-whu-40 | C2F-SemiCD | F1: 93.03 IoU: 86.97 KC: 92.74 OA: 99.45 Precision: 93.20 Recall: 92.86 |
| semi-supervised-change-detection-on-whu-5 | C2F-SemiCD | F1: 85.63 IoU: 74.87 KC: 85.04 OA: 98.87 Precision: 86.51 Recall: 84.77 |