Abstract
The image semantic segmentation algorithm required by edge intelligence needs to have high accuracy and real-time performance at the same time. In this paper, a new image segmentation network: Dense Aggregation based efficient neural network (DA-ENet) is proposed. The proposed network aggregates feature maps of each two adjacent stages of DA-ENet into an aggregation feature map, and uses shortcut connections connect each encoder stage feature map to the two decoder stages. In DA-Enet the linear propagation way of original ENet was changed. DA-Enet uses aggregations to realize feature fusion at different stages of the network and reduce the number of network parameters. DA-Enet uses shortcut connections to enhance the propagation and reuse of features. Aggregations and shortcut connections make DA-ENet a densely connected network. In densely connected network, each layer receives additional supervision from the loss function. Therefore, DA-Enet will be easier to train and has higher accuracy.The proposed DA-ENet was evaluated on three standard datasets: CamVid, CityScapes and SUN RGB-D. The experimental results demonstrate that DA-ENet has similar inference speed but higher segmentation accuracy to ENet.
Highlights
In the era of Internet of things, with the rapid development of technologies such as autonomous driving [1], [2], and robot navigation [3]–[5], the demand for real-time semantics segmentation technology is becoming more and more intense
We summarize the advantages of DA-Enet as follows: 1) The proposed Dense Aggregation based efficient neural network (DA-Efficient Neural Network (ENet)) retains ENet’s backbone network structure and bottleneck module in each stage of the network
EXPERIMENTS We evaluated the proposed DA-ENet on three different datasets: CamVid [39], Cityscapes [40] and SUNRGB-D
Summary
In the era of Internet of things, with the rapid development of technologies such as autonomous driving [1], [2], and robot navigation [3]–[5], the demand for real-time semantics segmentation technology is becoming more and more intense. We summarize the advantages of DA-Enet as follows: 1) The proposed DA-ENet retains ENet’s backbone network structure and bottleneck module in each stage of the network In this way, DA-ENet inherits the high efficiency and accuracy of ENet. 2) Aggregation and shortcut connections change the original linear feature propagation between ENet network modules. The accuracy of DA-ENet after adding shortcut connections is almost no better than that before adding these shortcut connections, but the frame processing time of DA-ENet after adding shortcut connections is about increased by 2 ms This is due to the injection of detail features into decoder stage 4 and 5, which is helpful to fine-tune segmentation details. We use 1×1 convolution to change the sizes of these concatenating feature maps into the input sizes required by stage 4 and stage 5
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
