Software para reconhecimento de espécies florestais a partir de imagens digitais de madeiras utilizando deep learning

Abstract

Classifying forest species is an essential process for the correct management of wood and forest control. After cutting off the trunk of the tree, many of the characteristics of the species are lost and identifying them becomes a much more difficult task. In this context, an anatomical analysis of the wood becomes necessary by specialists who know very well the cellular structures in each species. However, such methodology approaches few automated techniques, making it a delayed and error-prone activity. These factors undermine environmental control and decisionmaking. The use of computer vision is an alternative to automatic recognition, since it allows the development of intelligent models which, from images, are able to detect features and perform a final classification. There are some techniques that are state of the art in Digital Image Processing and Artificial Intelligence, such as, for example, convolutional neural networks. These networks are deep learning techniques that consists of creating models from raw data (images, audios, etc.). In this technique, the minimum preprocessing is used in the images and, in the training process, these models are able to learn the convolutional filters to be applied in feature extraction process. In this way, this technique abstracts a lot from the pre-processing complexity existing in traditional approaches. This research addresses some experiments using traditional techniques and experiments using convolutional neural networks (the focus of the research). Two original datasets were used, one including macroscopic images and the other including microscopic images, for which three models were created: scale recognition, species recognition from macroscopic images and species recognition from microscopic images. The best models provide 100% recognition rates for the scale dataset, 98.73% for the macroscopic datasetand 99.11% for the microscopic dataset, which they were better results than the ones obtained for the respective datasets in related researches. Finally, a web software was developed as a final product, using the three best models.