Light-Weight Edge Enhanced Network For On-Orbit Semantic Segmentation

ARTIFICIAL NEURAL NETWORKS AND MACHINE LEARNING - ICANN 2019: DEEP LEARNING, PT II(2019)

Cited 2|Views13
No score
Abstract
On-orbit semantic segmentation can produce the target image tile or image description to reduce the pressure on transmission resources of satellites. In this paper, we propose a fully convolutional network for on-orbit semantic segmentation, namely light-weight edge enhanced network (LEN). For the model to be pruned, we present a new model pruning strategy based on unsupervised clustering. The method is performed according to the l(1)-norm of each filter in the convolutional layer. And it effectively guides the pruning of filters and corresponding feature maps in a short time. In addition, the LEN uses a trainable edge enhanced module called enhanced domain transform to further optimize segmentation performance. The module fully exploits multi-level information of the object to generate the edge map and performs edgepreserving filtering on the coarse segmentation. Experimental results suggest that the models produce competitive results while containing only 1.53 M and 1.66 M parameters respectively on two public datasets: Inria Aerial Image Labeling Dataset and Massachusetts Buildings Dataset.
More
Translated text
Key words
Semantic segmentation, Model pruning, Enhanced domain transform
AI Read Science
Must-Reading Tree
Example
Generate MRT to find the research sequence of this paper
Chat Paper
Summary is being generated by the instructions you defined