Very Low Complexity Convolutional Neural Network for Generating Quadtree Structures

Publisher:
ARAA
Publication Type:
Conference Proceeding
Citation:
ACRA 2018 Website Proceedings, 2018, pp. 1 - 8
Issue Date:
2018-12-06
Full metadata record
In this paper, we present a Very Low Complexity Convolutional Neural Network (VLC-CNN) for the purpose of generating quadtree data structures for image segmentation. The use of quadtrees to encode images has applications including video encoding and robotic perception, with examples including the Coding Tree Unit in the High Efficiency Video Coding (HEVC) standard and Occupancy Grid Maps (OGM) as environment representations with variable grid-size. While some methods for determining quadtree structures include brute-force algorithms or heuristics, this paper describes the use of a Convolutional Neural Network (CNN) to predict the quadtree structure. CNNs traditionally require substantial computational and memory resources to operate, however, VLCCNN exploits downsampling and integer-only quantised arithmetic to achieve minimal complexity. Therefore, VLC-CNN’s minimal design makes it feasible for implementation in realtime or memory-constrained processing applications.
Please use this identifier to cite or link to this item: