Análise de viabilidade do uso de aprendizagem profunda para detecção de frutos de acerola em imagens.

Abstract

Ceará is the second-largest producer of acerolas in Brazil with 14.32% of the Brazilian production of the fruit, mostly, turned to export. The growing demand for green acerolas has contributed to boosting fruit production in some Brazilian states. Knowledge about fruit production in crops and their respective maturation stages can guide fruit growers in the management of labour resources, as well as support decisions for future planning. To develop an automated robotic harvesting system for acerola fruits, the main and most costly task to be developed is the detection of fruits. In this work, an acerola fruit detection module will be developed, involving the steps from the acquisition of the database to the detection of the fruit in images. To our best knowledge, there are no specific studies on the detection of acerola fruits through digital images. The main challenges encountered are in the acquisition and formation of the database, where through computational experiments, it was observed that the quality and quantity of information in the database cause significant impacts on the performance of the model. To develop the fruit detection module, the Faster-RCNN network architecture was used. To evaluate the model, the gls mAP metric was used, where it obtained 89.5% mAP with training data and 92.2% on the test data, in both cases the value was used from 0.5 to IoU.