Centre de documentation
scientifique

Line-based deep learning method for tree branch detection from digital images / Rodrigo L. S. Silva in International journal of applied Earth observation and geoinformation, vol 110 (June 2022)  

Public [article]  inInternational journal of applied Earth observation and geoinformation > vol 110 (June 2022) . - n° 102759 Titre : Line-based deep learning method for tree branch detection from digital images Type de document : Article/Communication Auteurs : Rodrigo L. S. Silva, Auteur ; José Marcato Junior, Auteur ; Laisa Almeida, Auteur ; et al., Auteur Année de publication : 2022 Article en page(s) : n° 102759 Note générale : bibliographie Langues : Anglais (eng) Descripteur : [Vedettes matières IGN] Traitement d'image optique [Termes IGN] apprentissage profond [Termes IGN] branche (arbre) [Termes IGN] classification par réseau neuronal convolutif [Termes IGN] détection d'objet [Termes IGN] données qualitatives [Termes IGN] estimation quantitative [Termes IGN] image à haute résolution [Termes IGN] ligne (géométrie) [Termes IGN] transformation de Hough Résumé : (auteur) Preventive maintenance of power lines, including cutting and pruning of tree branches, is essential to avoid interruptions in the energy supply. Automatic methods can support this risky task and also reduce time-consuming. Here, we propose a method in which the orientation and the grasping positions of tree branches are estimated. The proposed method firstly predicts the straight line (representing the tree branch extension) based on a convolutional neural network (CNN). Secondly, a Hough transform is applied to estimate the direction and position of the line. Finally, we estimate the grip point as the pixel point with the highest probability of belonging to the line. We generated a dataset based on internet searches and annotated 1868 images considering challenging scenarios with different tree branch shapes, capture devices, and environmental conditions. Ten-fold cross-validation was adopted, considering 90% for training and 10% for testing. We also assessed the method under corruptions (gaussian and shot) with different severity levels. The experimental analysis showed the effectiveness of the proposed method reporting F1-score of 96.78%. Our method outperformed state-of-the-art Deep Hough Transform (DHT) and Fully Convolutional Line Parsing (F-Clip). Numéro de notice : A2022-550 Affiliation des auteurs : non IGN Thématique : FORET/IMAGERIE Nature : Article DOI : 10.1016/j.jag.2022.102759 Date de publication en ligne : 09/05/2022 En ligne : https://doi.org/10.1016/j.jag.2022.102759 Format de la ressource électronique : URL article Permalink : https://documentation.ensg.eu/index.php?lvl=notice_display&id=101153 [article]