Proposta de bag-of-visual-words por meio de redes complexas

Abstract

The use of vigorous seeds is paramount to the success of the crop. Among the techniques of seed analysis that aim to identify seed batch vigor, the tetrazolium test stands out for the number of information provided. However, the tetrazolium test has the result strongly tied to the empiricism of the seed analysts. The present work aims to develop a methodology of automatic evaluation by means of images of the vigor of the soybean seeds submitted to the tetrazolium test, which serves as a check for the seed analyst, in order to increase the reliability and accelerate the results. It was used the Bag-of-Visual-Words method and conducted experiments with different detectors and descriptors of the literature in order to find the pipeline that presented the best cost-benefit in the categorization of seed vigor, obtaining 78.95% of average accuracy as the best result. In the present work it was also proposed the use of Complex Networks in the description stage of the Bag-of-Visual-Words method. The methodology proposed using only one threshold in the subgraph analysis step presented better results than all the thresholds, in addition to being more cost-effective than some methods of description in Bag-of-Visual- Words popular in the literature. These results identify the set of detectors and descriptors of the Bag-of-Visual-Words method that are more accurate and cost-effective for the vigor analysis of the seeds submitted to the tetrazolium test, besides the proposed merger of the Bag-of- Visual-Words and Complex Network.