Abstract

In this paper, we present an improved Efficient Neural Network (ENet) for semantic segmentation, and named the proposed network as Efficient Residual Neural Network (ERNet). The ERNet network contains two processing streams: one is pooling stream, which is used to obtain high-dimensional semantic information; the other is residual stream which is used to record low-dimensional boundary information. The ERNet has five stages, each stage contains several bottleneck modules. The output of each bottleneck in the ERNet network is fed into the residual stream. Starting from the second stage of ERNet, pooling stream and residual stream through concatenating are used as inputs for each down-sampling or up-sampling bottleneck. The identity mapping of residual stream shortens the distance between the near input and output terminals of each stage network in ERNet, alleviates the problem of vanishing gradient, strengthens the propagation of low-dimensional boundary features, and encourages feature reuse of low-dimensional boundary features. We tested ERNet on CamVid, Cityscape, and SUN RGB-D datasets. The segmentation speed of ERNet is close to that of ENet, but the segmentation accuracy is higher than that of ENet.

中文翻译：

---

用于语义分割的高效残差神经网络

摘要

在本文中，我们提出了一种用于语义分割的改进的高效神经网络（ENet），并将所提出的网络命名为高效残差神经网络（ERNet）。ERNet 网络包含两种处理流：一种是池化流，用于获取高维语义信息；另一个是残差流，用于记录低维边界信息。ERNet 有五个阶段，每个阶段包含几个瓶颈模块。ERNet 网络中每个瓶颈的输出都被输入到残差流中。从 ERNet 的第二阶段开始，通过连接将池化流和残差流用作每个下采样或上采样瓶颈的输入。残差流的恒等映射缩短了ERNet中每个阶段网络的近输入和输出终端之间的距离，缓解了梯度消失的问题，加强了低维边界特征的传播，鼓励了低维边界特征的特征重用. 我们在 CamVid、Cityscape 和 SUN RGB-D 数据集上测试了 ERNet。ERNet 的分割速度接近 ENet，但分割精度高于 ENet。