Aplicação de técnicas de aprendizado profundo para segmentação: um estudo de caso em folhas de café

Abstract

Disease incidence has always been a serious problem in coffee culture, capable of causing severe damage and significantly reducing production. In this sense, the early and correct identification of the symptoms caused by diseases is an important task to allow quick treatment and to mitigate the damage. One area of computing that can help alleviate this problem is deep learning, responsible for improving the state of art in several domains. However, the use of deep neural networks for this purpose still faces several challenges, mainly related to the complexity associated with the use of images in real circumstances, with variations in the general capture conditions or complex backgrounds. The application of segmentation techniques can contribute to a significant improvement of the results, since it enables the separation of the relevant aspects of the image and allows the network to focus on the right elements. Thus, this work proposed the application, evaluation and comparison of deep learning techniques for the segmentation of coffee leaves. Five segmentation models (U-Net, FPN, DeepLabv3+, CFNet and OCRNet) were tested with four different feature extractor networks (ResNet-50, ResNet- 152, DenseNet-121 and EfficientNet-B3). All trained models presented satisfactory performance, with IoU above 83% and F-score above 90%, showing them to be valid alternatives for the task at hand. The best results were obtained in the trainings that combined U-Net with DenseNet-121 (87.79% of IoU and 93.23% of F-score), U-Net with EfficientNet-B3 (87.10% of IoU and 92.89% of F-score), and FPN with DenseNet-121 (87.05% of IoU and 92.82% of F-score).