Aplicación de redes neuronales en el proceso de inspección de calidad del Aguacate Hass para exportación

Abstract

The Hass avocado is considered in recent years, one of the tropical fruits with the greatest prospect of growth in exports worldwide thanks to its multiple benefits and uses. The success of commercial exchange depends to a large extent on compliance with physical quality standards, which are verified at the plant through visual inspection by human operators who, one by one, severely identify compliance with their expertise and knowledge. However, since the process depends on human operators, it is susceptible to higher resource costs as a result of omissions in the verification of units, longer classification time and, in the event of incurring in breaches in the quality of the load, the suspension of the commercial exchange. For this reason, this research evaluates the application and use of three (3) supervised learning convolutional neural networks (shallow CNN, Google's Inception Learning and ResNet 50) as inspection and classification mechanisms by computer. This given that today, Artificial Intelligence, especially Neural Networks, are a useful and practical method to automate or improve tasks that were previously not possible through conventional technologies and now are, thanks to the analogy of human learning. In the proposed approach, in each of the architectures, the ReLU activation function, Batch Normalization, is used to reduce bias and increase information. Likewise, Glorot is used as an initialization scheme and precision as a network performance metric. Regarding the data bank, a photographic sample of 320 images of suitable units and 325 of units not suitable for export is made, which in order to strengthen the volume of the data were subjected to Data Augmentation, through Python and the well-known api Keras, achieving a total bank of 1,177 photos of suitable units and 1,273 photos of unsuitable units. Likewise, for the training, validation and test phases, the 60-40 technique was extracted. The results show that the proposed ResNet50 model achieves the classification of the units with 95% accuracy. Superior value with respect to simple CNN and Google's Inception Learning