Redes neurais convolucionais aplicadas na identificação de rachaduras em imagens de toras de eucalipto em campo

Abstract

Wood is a product widely used by human. Most of its stock comes from artificial forest stands promoting various industries such as civil construction and furniture. Although part of the stands presents good performances in terms of production, the sector still deals with material losses due to the projection of cracks in the wood. Efforts to minimize this problem mainly focus on obtaining trees with a genetic constitution that is less prone to cracking. Advances in genetic improvement for this character are reported, above all, for the evaluation methods. Traditional computer vision approaches are an excellent alternative for evaluating cracks. However, its inability to implement it in the field is the main bottleneck of the technique. Based on deep learning, field phenotyping approaches using computer vision show promise for field evaluation tasks. Thus, this work aimed to evaluate the ability of convolutional neural networks to identify cracks in logs from images collected in the field. For this purpose, images were used in the field and a controlled environment. The first group of images underwent class annotation processes (background, log, and crack) to generate masks. In contrast, masks previously made by other works were used for the second group. Synthetic images were generated from the controlled environment images, simulating the field environment from the segmentation and merging of the log-in field photos. In addition, several wood textures were combined to generalize the model. The set of images obtained, both synthetic and field, underwent database expansion processes so that the architectures have enough images for good training. In the end, 2,554 images were obtained, which were divided into sets of exclusive training and validation data in the proportion of 80% and 20% respectively. Three architectures of convolutional neural networks were evaluated, namely U-Net, FPN, and Linkinet. The models generated from the architectures were evaluated, and the best one was selected based on the best averages for insertion over union (IoU). The class estimates of the model chosen were visually compared with the expected masks from a set of 20 images. Finally, the actual and estimated area of the crack was calculated and compared using the chi-square test, coefficient of determination (R2), root mean squared error (RMSE), and the mean percentage absolute error (MAPE). With this work, it is observed that the U-Net model is configured as the best option among the architectures evaluated here for the phenotyping of cracks in logs in the field. The statistical equity of the estimates of the crack area concerning the real area and the speed in obtaining the results encourage its use in crack evaluation tasks in forest improvement programs, even though the model presents difficulties in dealing with small cracks. Future studies may focus on strategies to improve the detection of small cracks.