Scholarlink

Dry coffee beans are one of the main export products of the coffee productivity sector in Imbabura-Ecuador. The selection of these beans is done manually, which may lead to a decrease in product quality due to the limitations of the human factor, such as tiredness, the time needed, or different classification criteria. This work aims to detect morphological defects in dry coffee beans using convolutional neural networks (CNNs). The collected dataset consists of 3276 images manually labelled by an expert between good and bad beans. Two architectures were selected for experimentation: SSD-MobileNet v2 and SSD-Resnet50, using the Tensorflow object detection API. The results show an accuracy of 91.65% and 83.07%, respectively. The best model offers higher precision than the obtained, on average, by a manual selection (78.9%), with a light weight of 26.5 MB and an inference speed of 2.3 fps. These results show a possibility of improving it to reach a practical application for small and medium producers and associations.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Morphological Defects Classification in Coffee Beans Based on Convolutional Neural Networks

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Marcel Cevallos,
  • Lucía Sandoval-Pillajo,
  • Víctor Caranqui-Sánchez,
  • Cosme Ortega-Bustamante,
  • Marco Pusdá-Chulde,
  • Iván García-Santillán

摘要

Dry coffee beans are one of the main export products of the coffee productivity sector in Imbabura-Ecuador. The selection of these beans is done manually, which may lead to a decrease in product quality due to the limitations of the human factor, such as tiredness, the time needed, or different classification criteria. This work aims to detect morphological defects in dry coffee beans using convolutional neural networks (CNNs). The collected dataset consists of 3276 images manually labelled by an expert between good and bad beans. Two architectures were selected for experimentation: SSD-MobileNet v2 and SSD-Resnet50, using the Tensorflow object detection API. The results show an accuracy of 91.65% and 83.07%, respectively. The best model offers higher precision than the obtained, on average, by a manual selection (78.9%), with a light weight of 26.5 MB and an inference speed of 2.3 fps. These results show a possibility of improving it to reach a practical application for small and medium producers and associations.