Uma base de imagens de folhas de feijão e uma rede neural profunda para estimativa não-destrutiva de área foliar

Abstract

Leaves perform a fundamental role for the plant body doing photosynthesis and morphological characteristics (e.g., leaf area) associated with its surface are parameters that could help explain responses to various processes, such as climate change, ecological relationships, and agricultural productivity. Most of the existing methods for measuring leaf surface dimensions are expensive and often complicated. In addition, it commonly uses destructive approaches that make it impossible to monitor plant growth. In this context, a new database with annotated images was constructed for non-destructive estimation of bean leaf dimensions (area, width, length, and perimeter), based on a virtual reality marker added to the scene. The construction of the database involved a process of planting, image acquisition, leaf picking, manual measurement, semi-automatic segmentation, and marker pose estimation. Furthermore, a new deep neural network was developed that receives one input image containing a salient leaf accompanied by a marker and provides relative leaf area by comparing the proportions of both objects in the image. The proposed method is based on the architecture of a semantic segmentation neural network. The main hypothesis is that is possible to adapt a convolutional neural network to regress image pixels area. Thus, a new decoder module is proposed for the network, used to remap image representation on relative area estimation of the objects of interest, leaf and marker. The model presented is composed of one encoder and two decoders, which estimate the image segmentation and the pixels area of the objects of interest. A way of calculating the loss of the decoder and selection criteria for the best model are also defined. For proposal viability determination, an extensive quantitative and qualitative analysis is performed with the model’s predictions on 1033 images of 90 different leaves. Results indicate that the model is capable of learning to estimate the area of the objects of interest with only one input image.