Fully convolutional siamese networks for change detection

arXiv:1810.08462v1  [cs.CV]  19 Oct 2018

论文地址：https://arxiv.org/abs/1810.08462v1

部分代码：https://github.com/rcdaudt/fully_convolutional_change_detection

ABSTRACT

提出三种全卷积的结构用于已配准的图像对的变化检测，可检测RGB和多光谱的图像。从头的监督训练，比related systems 快500倍。

1. INTRODUCTION

变化检测缺少大型的带注释的数据集，但是，还是有一些能获得的pixelwise的数据集。

2. RELATEDWORK
指出迁移学习方法的局限性：

 This is limiting in many ways, as it assumes similarities between these datasets and the relevant change detection data. For example, most of the networks have been trained on RGB images, and cannot be transfered for SAR or multi-spectral images.

 These methods also avoid end-to-end training,which tends to have better results for successfully trained systems.

3. PROPOSED APPROACH
所提出的结构是从[3]《Urban change detection for multispectral earth observation using convolutional neural networks》（论文 | 笔记）演变过来的。没有使用[3]中 patch-based 的方法，这样提高了预测阶段的速度和精度。

我们使用了用于U-Net的跳过连接概念，这样做的动机是用在网络较早的层中存在的空间细节来补充编码之后的更抽象和全局化的信息，以在输出图像中产生具有精确边界的准确类预测。

第一种结构是 Fully Convolutional Early Fusion (FC-EF)， The FC-EF (Fig. 1(a)) contains therefore only four max pooling and four upsampling layers, instead of the five present in the UNet model. The layers in FC-EF are also shallower than their U-Net equivalents.与在[3]中patch based的EF一样，都是先把图像对先拼接起来。

另外两种都是孪生结构，只是skip connections 的方式不同。

第二种更直观的方法是在解码步骤中concat两个 skip connections ，每个 skip connections 来自一个encoding stream。 这种方法被称为Siamese-Concatenation（FC-Siam-conc，图1(b)）。有图可以看出在decoder的过程中，每次是三个feature map concat到一起。

第三种方法，是concat了它们差的绝对值。 这种方法被称为Fully Convolutional Siamese - Difference（FC-Siam-diff，图1(c)）。

4. EXPERIMENTS

用两个数据集来评估，the Onera Satellite Change Detection dataset [3] (OSCD), 和 the Air Change Dataset [4] (AC)。AC中是RGB图像，OSCD是多光谱图像。

这两个类别( change and no change )的权重与每一个类别的样本数量成反比。 The available data was augmented by using all possible flips and rotations multiple of 90 degrees to the training patches. Dropout was used to help avoid overfitting during training. 

In these tests, the FC-Siam-diff architecture seems to be the most suited for change detection,followed closely by FC-EF.

原因： 

 First, fully convolutional networks were developed with the express purpose of dealing with dense prediction problems, such as CD.

Second, the Siamese architecture imbues into the system an explicit comparison between two images.

Finally, the difference skip connections also explicitly guides the network to compare the differences between the images, in other words, to detect the changes between the two images