Title: Computer image analysis for various shading factors segmentation in forest canopy using convolutional neural networks

Authors: Liangkuan Zhu; Jingyu Wang; Kexin Li

Addresses: School of Mechanical and Electrical Engineering, Northeast Forestry University, Harbin 150040, Heilongjiang, China ' School of Mechanical and Electrical Engineering, Northeast Forestry University, Harbin 150040, Heilongjiang, China ' School of Mechanical and Electrical Engineering, Northeast Forestry University, Harbin 150040, Heilongjiang, China

Abstract: Determination of images of the various parts in the forest canopy is critical because it reflects a variety of parameters for plant population growth in forest ecosystems. This study presents the use of deep learning for the detection of various shading factors in hemispherical photographs of the forest canopy. First, a forest canopy hemispherical photographs dataset that can be used for the research of related algorithms is constructed. Based on FCN, the dilated convolution layers and multi-scale feature fusion are used to improve the accuracy of forest canopy image segmentation. Furthermore, the Conditional Random Field (CRF) is adopted to optimise the results. Finally, experiments show that this method can achieve automatic segmentation of the sky, leaves, and trunks of forest canopy images. Compared with the FCN model, the average pixel accuracy of the improved FCN model is improved by 9.11%, and it has good robustness.

Keywords: hemispherical photographs; image segmentation; full convolutional neural network; dilated convolution; multi-feature fusion; conditional random field.

DOI: 10.1504/IJCAT.2020.10034802

International Journal of Computer Applications in Technology, 2020 Vol.64 No.4, pp.415 - 428

Received: 23 Mar 2020
Accepted: 29 Mar 2020

Published online: 28 Jan 2021 *

Full-text access for editors Access for subscribers Purchase this article Comment on this article