Segmentação e contagem de troncos de madeira utilizando deep learning e processamento de imagens

Abstract

Counting objects in images is a pattern recognition problem that focus on identifying an element to determine its incidence. Approached in the literature as Visual Object Counting (VOC), such application can be very precise. In this work we propose a methodology to count wood logs. First, wood logs are segmented from the image background and their incidence is evaluated in a second step. This first segmentation step is created using the Pix2Pix framework that uses Conditional Generative Adversarial Networks (CGANs), which implements pixel-by-pixel segmentation. In the second step, in order to count the segmented clusters, we use and compare the following approaches: Hough Transform, Connected Components and Morphological Reconstruction. The average accuracy of the segmentation exceeds 89%, while the average amount of correctly identified trunks is over 97%.