CyclicNet: an alternately updated network for semantic segmentation
- PDF / 1,991,358 Bytes
- 15 Pages / 439.642 x 666.49 pts Page_size
- 47 Downloads / 206 Views
CyclicNet: an alternately updated network for semantic segmentation Guixian Wu1 · Yuancheng Li1 Received: 25 May 2020 / Revised: 6 August 2020 / Accepted: 2 September 2020 / © Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract In recent years, with the continuous breakthrough in deep learning, convolutional neural networks (CNNs) have shown great potential for semantic segmentation. CNNs achieve better results by deepening or widening the network, but they increase the utilization rate of computing resources and even have the phenomenon of the vanishing gradient. A new convolutional neural network architecture with alternately updated clique (CliqueNet) can get a deeper network and improves the utilization of network features. In order to maximize the transmission of semantic information, this paper introduces the clique block of CliqueNet and proposes a new fully convolutional network based on the encoder-decoder structure, which calls the CyclicNet, an alternately updated network for semantic segmentation. Besides, the long skip connections and short skip connections are added in the network to avoid the vanishing gradient. The experiment was conducted on the CamVid and Cityscapes. Comparing it with the current advanced architectures shows that CyclicNet can maximize information flow and achieve the most superior results. Keywords Semantic segmentation · Convolutional neural networks · Fully convolutional networks · Image understanding
1 Introduction With the rapid development of CNNs, deep learning has achieved great success in computer vision. The proposed of AlexNet [14] established the dominant position of CNNs in computer vision. Since then, the CNNs based on the neural network have been put forward one after another, among which the representative ones are VGG-Net [27], GoogLeNet [29], Yuancheng Li
[email protected] Guixian Wu [email protected] 1
School of Control and Computer Engineering, North China Electric Power University, Beijing, China
Multimedia Tools and Applications
ResNet [11], DenseNet [13], CliqueNet [38], etc. They have made breakthroughs again and again in ImageNet competition. The biggest characteristic of these CNNs is that the structures are deeper, more accurate, and lighter, even more accurate than human recognition in some areas. The convolution layer on CNNs can effectively capture the local features in the image, and many of these modules are nested in a layered way, which makes CNNs extract larger structures. Recently, CNNs have been widely used in three core areas of image classification, semantic segmentation, and object detection in computer vision [18, 31, 32]. Semantic segmentation refers to identifying objects of the same pixel class in the image and classifying different object classes [4]. It has been widely applied in medical image diagnosis [6, 22, 40], pedestrian detection [23, 24, 37], video object segmentation [8, 20, 34] and other fields [3, 5, 10, 28], which has great research significance and value. Semantic segmentation is based on t
Data Loading...
