Sistema de visão computacional para a contagem de sementes de soja

Abstract

Seed counting is an important step for the correct pricing and further commercialization of the seeds. The traditional methods of seed counting are based on the weigh or volume measurement, which is prone to errors. Thus, such traditional seed counting methods can be improved in order to provide a larger accuracy. This work presents a computer vision system for the counting of soybean seeds. The developed system is based on a conveyor belt, a RGB camera and open source image processing software libraries. Main components of the developed algorithm are: noise removing, HSV-based color segmentation, morphological operations, watershed and labeling. Results shows that the proposed system presents a mean absolute error of 5.32%. This error is adequate for the current stage of this work, where a first version of the system for the counting of soybean seeds. However, it is important to mention that improvements are required for the eventual adoption of the developed system in industrial applications, since a better accuracy is necessary.