Visão computacional associada a deep learning na avaliação da qualidade de sementes

Abstract

The seed is an indispensable input for agribusiness, carrying with it all the research and development of genetic improvement. In this context, the use of technologies is a differentiator, gaining increasing importance in this vital sector of the economy. Computer vision (CV) associated with image-based detection systems is of great importance in determining agri-food quality, providing reliability, precision, and speed, and eliminating instability and inconsistency through human intervention. Deep learning techniques associated with digital image processing can offer numerous benefits, including the development of computational models for decision-making, enabling both large and small farmers to optimize decision-making at various stages of production. Therefore, the objective of this work is to use CV associated with deep learning in seed quality assessment, aiming to eliminate subjectivity and optimize the analysis process. The study was conducted at the Seed Analysis Laboratory of the Department of Crop Science at the Federal University of Ceará (UFC), in collaboration with the UFC’s Department of Computer Science and the University of Cagliari in Italy. Ten batches of cowpea seeds (Vigna unguiculata (L.) Walp.) were used. The research focused on creating an artificial vision system for analyzing and classifying seeds after performing the tetrazolium test. The software was designed to allow image analysis, generate results based on light crimson red, dark crimson red, and milky white colors, and automatically classify the seeds according to current literature on vigor and viability. In addition, images of untreated seeds were analyzed using the Seeds Analysis plugin, part of the ImageJ software, which was validated in Italy to discriminate morphocolorimetric characteristics. The seed vigor was evaluated using traditional tests for characterization, comparison, and validation with automated techniques. All methodologies addressed in the thesis demonstrated high efficiency for automated analysis, eliminating subjectivity and increasing precision and efficiency by recognizing hidden patterns, correlating data, grouping, classifying, and providing this knowledge for updating according to innovations in the seed sector. The development of an artificial vision system proved to be an effective solution to overcome the limitations of traditional methods. Furthermore, the use of tools such as the Seeds Analysis plugin validated the potential of these technologies in identifying and quantifying morphocolorimetric patterns in seeds, significantly contributing to advances in the seed sector. Thus, this work not only highlights the feasibility and relevance of using these technological innovations but also paves the way for future updates and improvements in the field of seed analysis.