LinkNet: Exploiting Encoder Representations for Efficient Semantic Segmentation
https://arxiv.org/abs/1707.03718
Torch7 code: https://github.com/e-lab/LinkNet
本文主要侧重语义分割的速度问题,算法思路类似 U-Net,引入了 residual blocks
LinkNet Architecture
conv means convolution 和 full-conv means full convolution
/2 denotes downsampling by a factor of 2 which is achieved by performing strided convolution 降采样2倍
∗2 means upsampling by a factor of 2 上采样2倍
输入图像首先使用一个 7*7的卷积核卷积,再用一个 步长为2的 3*3 spatial max-pooling
encoder-block 如下:
decoder-block 如下:
各个模块特征图数量:
首先来看看速度如何:
模型大小和运算量:
再来看看性能