COMPUTER VISION BASED APPROACHES FOR BEAN DEFECTS DETECTION

Resumo

In this work are proposed computer vision approaches to detect three of the main defects
found in beans: broken, bored by insect (Acanthoscelides obtectus) and moldy. In addition, we describe
a fast and robust segmentation step that is combined with the proposed approaches to compose a
computer vision system (CVS) applicable to the Brazilian beans quality inspection process, to determine
the type of the product. The proposed approaches constitute an important practical contribution since,
although there are some papers in the literature addressing visual inspection of beans, none of them deals
with defects. In the conducted experiments a low-cost equipment, composed by a table made in
structural aluminum, a conveyor belt and an image acquisition chamber, was used to simulate the
characteristics of an industrial environment. The CVS evaluation was performed in two modes: offline
and online. In the offline mode, a database composed by 120 images of bean samples containing grains
of different classes and with different defects was employed, while in the online mode the grains
contained in a batch were spilled continuously in the conveyor belt of the equipment for the proposed
CVS to perform the tasks of segmentation and detection of defects. In the experiments the CVS was
able to process an image of 1280×720 pixels in approximately 2 s, with average hit rates of 99.61%
(offline) and 97.78% (online) in segmentation, and 90.00% (offline) and 85.00% (online) in detecting
defects.