Processamento digital de imagens na identificação de variedades de soja

Abstract

This work presents an alternative to the characterization of soybean cultivars in order to reduce spending typically aggregated in this task. For this, it was used a set of digital image processing techniques and artificial neural networks, as suggested in Padilha (2007). The fundamental difference of this work from the work of Padilha (2007) is the use of grain color as the main factor for the recognition of varieties. The use of color aims to improve the process. In this sense, it was used a case study for the following soybean cultivars: AL 83, BRS 133, BRS 184, BRS 214, CD 205, CD 206, CD 215, EMB 48, MERCEDES 70, MSOY 5826, NK 8350, RS 10 and MSOY 8000 RR. The images were supplied by Padilha (2007) and they were pre-processed until individual grains could be precisely located in the images. Then, the grain RGB coefficients in the images provided quantitative data needed to feed a neural network. Finally, several architectures of feed forward single hidden layer were trained by the back propagation algorithm and there was selected the model which shown better generalization ability to solve the problem.