Explorando técnicas de deep learning para segmentação e contagem de objetos em imagens: aplicação em pilhas de toras de madeira

Abstract

Accurate counting of timber logs constitutes a fundamental challenge in forest management. Traditionally manual processes present inconsistencies, operator fatigue, and precision limitations that result in substantial economic losses. Considering that the timber industry in Brazil annually produces 158,283,790 cubic meters of logs, with 69.70% coming from eucalyptus according to IBGE data, the automation of these processes represents strategic economic potential and a necessity for sector modernization. This dissertation explores deep learning techniques for automation of counting through segmentation of eucalyptus trunks present in images of wood piles. As the methodology proposed from this research, Conditional Generative Adversarial Networks (CGANs) are employed, representing a pioneering application of the Pix2Pix framework, traditionally used for image-to-image translation, systematically adapting it for object quantification in contexts characterized by extreme lighting variations, occlusions, and irregularities of natural materials. The developed hybrid methodology combines deep learning with mathematical morphology techniques, implementing 15 iterations of morphological erosion as an optimized algorithmic solution for handling intersections between segmented objects. A proprietary dataset was developed with carefully cataloged eucalyptus logs, encompassing systematic variations in light intensity, diversified capture angles, and adverse environmental conditions. Additionally, the Intersection Sensitive Score (ISS) was proposed, a specific index for quantitative quality evaluation in counting tasks considering intersection challenges in segmentation. Rigorous experimental validation compared three distinct neural network architectures and three counting algorithms, demonstrating consistent superiority of Pix2Pix and Connected Components. The results achieved segmentation precision of 89.6%, F1 Score of 91.1%, Kappa of 86.8%, IoU of 84.2%, and counting accuracy exceeding 90%. The research materialized into mobile software that integrates algorithmic automation with interactive interface for manual refinement. The solution allows dynamic adaptation to field variabilities through adjustable threshold fields. The software implementation exclusively uses conventional RGB cameras from standard mobile devices, democratizing access to advanced technology without dependence on specialized equipment such as LiDAR scanners. The general methodology establishes a hybrid paradigm that combines computational efficiency with human expertise, offering a balanced solution between algorithmic precision and operational viability for sustainable modernization of the Brazilian forestry sector.