Sistema de visão computacional para inspeção da qualidade de grãos de feijão

Abstract

The visual properties of many agricultural products, including bean grains, are important factors in determining their market prices and assisting consumer choice. Basically, the visual inspection of Brazilian bean quality is done manually following the operational procedures established by the Ministry of Agriculture, Livestock and Supply, which instruct how to frame the beans in group (according to botanical species), class (based on the color mixture of the skins) and type (the summary of defects found in the inspected sample. Manual quality inspection processes are usually subject to problems such as the high cost and difficulty of standardizing the results. In this context, it is important to use computational systems to support such processes in order to reduce operational costs and standardize results, generating a competitive differential for companies. In this work was developed a computer vision system to inspect beans quality (class and type determination), called SIVQUAF, composed of a set of software and hardware. For software design, computational approaches were proposed for segmentation, classification and defects detection. The hardware consists of an equipment developed with low-cost electromechanical materials, such as a table made of structural aluminum that includes an image acquisition chamber, servo motor and grain separation mechanism. Experiments were performed with SIVQUAF in two modes: individualized sample and continuous. For the first mode, we used a database composed of 270 images of bean samples, with different mixtures and defects, that was acquired with the use of the developed equipment. In the continuous (or online) mode, the beans contained in a batch, for example a bag of 1 kg, are spilled continuously on the conveyor belt for the system to perform the inspection, similar to what occurs in the food industry. These experiments demonstrated the feasibility of SIVQUAF to operate in continuous mode, since it is capable of processing an image of 1280×720 pixels in approximately 1.0 s, with success rates of 98.0% in segmentation, 99.0% in classification and more than 80.0% in defects detection.