Abstract

Semantic segmentation can be considered as a per-pixel localization and classification problem, which gives a meaningful label to each pixel in an input image. Deep convolutional neural networks have made extremely successful in semantic segmentation in recent years. However, some challenges still exist. The first challenge task is that most current networks are complex and it is hard to deploy these models on mobile devices because of the limitation of computational cost and memory. Getting more contextual information from downsampled feature maps is another challenging task. To this end, we propose an asymmetric depthwise separable convolution network (ADSCNet) which is a lightweight neural network for real-time semantic segmentation. To facilitating information propagation, Dense Dilated Convolution Connections (DDCC), which connects a set of dilated convolutional layers in a dense way, is introduced in the network. Pooling operation is inserted before ADSCNet unit to cover more contextual information in prediction. Extensive experimental results validate the superior performance of our proposed method compared with other network architectures. Our approach achieves mean intersection over union (mIOU) of 67.5% on Cityscapes dataset at 76.9 frames per second.

中文翻译：

---

ADSCNet：用于实时语义分段的非对称深度可分离卷积

摘要

语义分割可以看作是每个像素的定位和分类问题，它为输入图像中的每个像素提供了有意义的标签。近年来，深度卷积神经网络在语义分割方面取得了极大的成功。但是，仍然存在一些挑战。第一个挑战任务是当前大多数网络都很复杂，并且由于计算成本和内存的限制，很难在移动设备上部署这些模型。从降采样的特征图中获取更多上下文信息是另一项艰巨的任务。为此，我们提出了一种非对称深度可分离卷积网络（ADSCNet），它是用于实时语义分段的轻量级神经网络。为了促进信息传播，密集膨胀卷积连接（DDCC）网络中引入了以密集方式连接一组扩张的卷积层的方法。在ADSCNet单元之前插入池化操作，以覆盖预测中的更多上下文信息。大量的实验结果验证了我们提出的方法与其他网络体系结构相比的优越性能。我们的方法实现了67.5的平均联合交叉（mIOU）在Cityscapes数据集上的百分比为每秒76.9帧。