DeconvNet算法的简介(论文介绍)
DeconvNet网络架构，是由Convolution network、Deconvolution network两种架构组成。
Convolution network：feature extractor，采用VGG-16提取特征；
Deconvolution network：shape generator，通过上采样，计算像素的类别得分图。
Abstract
We propose a novel semantic segmentation algorithm by learning a deconvolution network. We learn the network on top of the convolutional layers adopted from VGG 16- layer net. The deconvolution network is composed of deconvolution and unpooling layers, which identify pixel-wise class labels and predict segmentation masks. We apply the trained network to each proposal in an input image, and construct the final semantic segmentation map by combining the results from all proposals in a simple manner. The proposed algorithm mitigates the limitations of the existing methods based on fully convolutional networks by integrating deep deconvolution network and proposal-wise prediction; our segmentation method typically identifies detailed structures and handles objects in multiple scales naturally. Our network demonstrates outstanding performance in PASCAL VOC 2012 dataset, and we achieve the best accuracy (72.5%) among the methods trained with no external data through ensemble with the fully convolutional network.
本文提出了一种新的基于反卷积网络的语义分割算法。我们学习了VGG 16层网在卷积层之上的网络。反卷积网络由反褶积层和反池层组成，它们识别像素级标签并预测分割掩码。我们将训练好的网络应用于输入图像中的每个提案，并将所有提案的结果以一种简单的方式结合起来，构造出最终的语义分割图。该算法将深度反卷积网络与建议预测相结合，克服了现有全卷积网络方法的局限性;我们的分割方法通常识别详细的结构和处理对象在多个尺度自然。我们的网络在PASCAL VOC 2012数据集中表现出色，通过全卷积网络集成，在没有外部数据训练的方法中，我们的准确率最高(72.5%)。
Conclusion
We proposed a novel semantic segmentation algorithm by learning a deconvolution network. The proposed deconvolution network is suitable to generate dense and pre-cise object segmentation masks since coarse-to-fine structures of an object is reconstructed progressively through a sequence of deconvolution operations. Our algorithm based on instance-wise prediction is advantageous to handle object scale variations by eliminating the limitation of fixed-size receptive field in the fully convolutional network. We further proposed an ensemble approach, which combines the outputs of the proposed algorithm and FCNbased method, and achieved substantially better performance thanks to complementary characteristics of both algorithms. Our network demonstrated the state-of-the-art performance in PASCAL VOC 2012 segmentation benchmark among the methods trained with no external data.
本文提出了一种新的基于反卷积网络的语义分割算法。该反褶积网络通过一系列的反卷积操作，逐步重构出由粗到细的目标结构，适用于生成密集的预分割掩码。我们的基于实例预测的算法消除了全卷积网络中固定大小接受域的限制，有利于处理对象尺度变化。我们进一步提出了一种集成方法，将所提算法的输出与基于FCN的方法相结合，由于两种算法的互补特性，取得了较好的性能。在没有外部数据训练的方法中，我们的网络在PASCAL VOC 2012分割基准测试中展示了最先进的性能。
论文
Hyeonwoo Noh, SeunghoonHong, BohyungHan.
Learning deconvolution network for semantic segmentation, ICLR, 2015.
https://arxiv.org/abs/1505.04366
