Desarrollo de un prototipo de bajo costo para la identificación y clasificación de huevos de gallina con un sistema inteligente de aprendizaje profundo

Abstract

The evolution of technology and digital infrastructure has made it possible for artificial intelligence (AI) to be used as effective tools to solve industrial problems, allowing better optimization of resources. This study arises from the need to classify chicken eggs in local poultry farms, a task that is normally carried out by an inspector, but which can be exposed to errors or subjectivities due to fatigue, considering that humans maintain between 60% to 90% probability of error compared to 3% for machines. The main objective is to develop a prototype application that uses deep neural networks, executed on NVIDIA Jetson Orin Nano hardware for efficient automatic classification of chicken eggs. A quantitative methodology is applied, of the quasi-experimental type by operating variables such as number of eggs, images, percentages and classes; The type of study is empirical as it is a systematic and iterative process encompassing data, training scheme, programming, evaluation and testing of the model. The CRISP-DM methodology was applied to understand the business along with the prototype requirements. The population is all the eggs processed in one hour (36,000) and the sample is the eggs that pass through the conveyor belt in two minutes, around 1,200, taking 200 frames as a basis for the learning model. The results highlight that the level of precision when detecting/classifying chicken eggs into healthy (green), dirty (red), broken (yellow) and rough (blue) is greater than 95%, achieving a performance of 99% effectiveness, meeting the objective of the study. The prototype developed on NVIDIA hardware supports the You Only Look Once (Yolo V8n) neural network used to view and classify eggs with a low-cost budget in contrast to industrial solutions in between; becoming around 10% of a commercial machine capable of carrying out the same process; Furthermore, it stands out for its scalability, adaptation and low energy consumption; Therefore, it is a practical solution that satisfies the design requirements, demonstrating through the discussion that similar results are obtained in computer vision research when using Yolo to classify objects and even study movement patterns in living beings.