卫星图像分割--Effective Use of Dilated Convolutions for Segmenting Small Object Instances

Effective Use of Dilated Convolutions for Segmenting Small Object Instances in Remote Sensing Imagery https://arxiv.org/abs/1709.00179

dilated convolutions 是一个不错的方法，它可以保持分辨率不变，但是目前dilated convolutions 使用方法不能很好的分割 小目标， aggressively increasing dilation factors fails to aggregate local features of small objects. This means that whereas increasing dilation factors is important in terms of resolution and context, it can be detrimental to small objects. This is especially undesirable for remote sensing scenario.

We solve this problem by simply going against the tide—decreasingly dilated convolutions.

Overview of the proposed network architecture

3.3. Local feature extraction module LFE module 主要是解决 front-end module 的问题。大量使用 dilated convolution 造成了两个问题 (1) spatial consistency between neighboring units becomes weak 相邻神经元直接的空间联系变弱 (2) local structure cannot be extracted in higher layer. 在后面的网络层提取不到图像中的局部结构信息

Problem on spatial inconsistency： We can see that information pyramids of two adjacent units do not overlap due to the sparse connections of the dilated kernels 从上图我们可以看出因为 dilated kernels 中的稀疏连接，导致相邻神经元没有联系 In the case of the dilation factor of 2, two neighboring units have non-overlap information pyramids, and as we increase the dilation factor, number of neighboring units which have non-overlap information pyramids grows larger. 当我们增加 dilation factor 时， 没有联系的神经元空隙会变大。

Problem on local structure extraction: 上图的右边显示，网络前面层中间神经元的不联系导致网络后面层神经元之间不联系，局部结构信息的丢失 information pyramids do not overlap for two adjacent units in bottom most layer. All units in top most layer receive information from either of the two units, but not both. This means that all units in top most layer are unaware of local structure inside the two units.

Local feature extraction module: 解决的方法就是 先 increasing dilation factor，再 decreasing dilation factor

3.4. Post-processing 直接用二值化得到分割结果

4 Experiments Toyota City Dataset

Relative improvements

Massachusetts Buildings Dataset

Vaihingen Dataset

Massachusetts Buildings Dataset 分割结果图

0 条评论

相关文章

682

9687

3877

3146

1422

吊炸天的CNNs，这是我见过最详尽的图解！（下）

1855

上个月发布了四篇文章，主要讲了深度学习中的“hello world”----mnist图像识别，以及卷积神经网络的原理详解，包括基本原理、自己手写CNN和p...

3985

1140

6.5K3

Accuracy

$A_{cc}=\frac{1}{n}\sum{I(\hat{y}_i=y_i)}$

581