A Deep Learning-based Approach for Tree Trunk Segmentation

• Autor(a) principal: Jodas, Danilo Samuel
• Data de Publicação: 2021
• Outros Autores: Brazolin, Sergio, Yojo, Takashi, De Lima, Reinaldo Araujo, Velasco, Giuliana Del Nero, Machado, Aline Ribeiro, Papa, Joao Paulo [UNESP]
• Tipo de documento: Artigo de conferência
• Idioma: eng
• Título da fonte: Repositório Institucional da UNESP
• Texto Completo: http://dx.doi.org/10.1109/SIBGRAPI54419.2021.00057; http://hdl.handle.net/11449/234109
• topic: convolutional neural networks; Deep learning; image processing; semantic segmentation; urban forest

Recently, the real-time monitoring of the urban ecosystem has raised the attention of many municipal forestry management services. The proper maintenance of trees is seen as crucial to guarantee the quality and safety of the streetscape. However, the current analysis still involves the time-consuming fieldwork conducted for extracting the measurements of each part of the tree, including the angle and diameter of the trunk, to cite a few. Therefore, real-time monitoring is thoroughly necessary for the rapid identification of the constituent parts of the trees in images of the urban environment and the automatic estimation of their physical measures. This paper presents a method to segment the tree trunks in photographs of the municipal regions. To accomplish such a task, we introduce a semantic segmentation convolutional neural network architecture that incorporates a depthwise residual block to the well-known U-Net model to reduce the parameters required to create the network. Then, we perform a post-processing step to refine the segmented regions by removing the additional binary areas not related to the tree trunk. Lastly, the proposed method also extracts the central line of the identified region for future computation of the trunk measurements. Compared with the original U-Net architecture, the obtained results confirm the robustness of the proposed approaches, including similar evaluation metrics and the significant reduction of the network size.