Uma abordagem de visão computacional por meio de redes convolucionais para o reconhecimento de imagens de sementes de feijão

Abstract

The present study aims to apply Data Science techniques for the classification and recognition of bean varieties (Phaseolus vulgaris) using Convolutional Neural Networks (CNN) and Artificial Intelligence, leveraging architectures such as VGG 16 and 19, ResNet 50, and MobileNetV2. It addresses the challenge faced by farmers and scientists in identifying and distinguishing bean varieties, whose characteristics may be subtle and visually complex, leading to time-consuming and error-prone manual processes. To overcome these limitations, Deep Learning techniques, like Convolutional Neural Networks, and Artificial Intelligence are explored to create an efficient and accurate computational model for automated bean variety classification. The methodology involves collecting a dataset of bean seeds of various species, organizing it into a database, applying preprocessing and data augmentation techniques to optimize data quality and quantity for model training. The classification system implementation utilizes the architecture of Convolutional Neural Networks, with adjustments and optimizations for the specific task of bean species recognition. Model evaluation, using performance metrics like accuracy, achieved 72% accuracy for original images and 61% for images filtered by a Gabor filter, indicating the need to expand the databases for a real increase in efficiency.