FCN

卷积化，上采样，skip

摘要

Our key insight is to build “fully convolutional” networks that take input of arbitrary size and produce correspondingly-sized output with efficient inference and learning.
（建立“完全卷积”网络，这种网络接受任意大小的输入，并通过有效的推理和学习产生相应大小的输出。）

We adapt contemporary classification networks (AlexNet , the VGG net , and GoogLeNet ) into fully convolutional networks and transfer their learned representations by fine-tuning to the segmentation task.
（我们将分类网络适应到完全卷积网络，并通过微调将它们的学习表示转移到分割任务中。）

We then define a novel architecture that combines semantic information from a deep, coarse layer with appearance information from a shallow, fine layer to produce accurate and detailed segmentations.
（定义了一种体系结构，它将来自深层、粗略层的语义信息与来自浅层、细层的外观信息相结合，以产生准确而详细的分割）

1引言

第一个端到端的训练工作
将分类网络微调以进行分割工作
跳连接–skip

Semantic segmentation faces an inherent tension between semantics and location: **global information resolves what while local information resolves where.（全局信息决定什么，而局部信息决定哪里）

2相关工作

we adapt and extend deep classification architectures, using image classification as su- pervised pre-training, and fine-tune fully convolutionally to learn simply and efficiently from whole image inputs and whole image ground thruths.
（我们采用和扩展了深度分类体系结构，使用图像分类作为有监督的预训练，并进行充分卷积微调，以简单有效地从全图像输入和全图像ground thruths中学习）

3全卷积网络 Fully convolutional networks

3.1Adapting classifiers for dense prediction（自适应分类器）
3.2Shift-and-stitch is filter rarefaction
3.3Upsampling is backwards strided convolution
3.4Patchwise training is loss sampling

4架构–Segmentation Architecture

将粗略的高层信息与精细的低层信息相结合。 。仅显示池化和预测层；省略中间卷积层(包括我们转换的完全连接的层)。实线(FCN-32S)：我们的单流网络，在单个步骤中向上采样将32个预测跨步返回到像素。虚线(FCN-16s)：组合来自最后一层和pool4层的预测，步幅16，让我们的网络预测更精细的细节，同时保留高级语义信息。虚线(FCN-8S)：来自pool3的更多预测(步幅为8)提供了更高的精度

4.1 From classifier to dense FCN

convolutionalizing proven classification architectures （卷积化之前的构架）

4.2 Combining what and where

（skip连接）
combines layers of the feature hierarchy and refines the spatial precision of the output.（它结合了特征层次的各层，并细化了输出的空间精度。）
Combining fine layers and coarse layers lets the model make local predictions that respect global structure.（通过将细层和粗层相结合，模型可以做出尊重全局结构的局部预测）

4.3Experimental framework

5结果

6结论

Fully convolutional networks are a rich class of models, of which modern classification convnets are a special case. Recognizing this, extending these classification nets to segmentation, and improving the architecture with multi-resolution layer combinations dramatically improves the state-of-the-art, while simultaneously simplifying and speeding up learning and inference.