Clasificación de residuos mediante una aplicación android con un modelo entrenado para el aprendizaje en niños de primaria

Abstract

This research addresses the problem of incorrect waste classification through the development of "MiniEco", an Android mobile application integrated with artificial intelligence for primary school students at the Juan León Mera La Salle Educational Unit. The Mobile-D methodology was used for software development and CRISP-DM for artificial intelligence. After initially training a ResNet50 architecture with TrashNet, it failed in real-time due to unvaried backgrounds and camera frame processing issues. Therefore, the system was optimized using a more diverse, real world dataset and selecting the EfficientNetV2S model. This model was converted to TensorFlow Lite to ensure 100 percent offline operation, achieving a laboratory accuracy of 94.50 percent with a final application size of 36 MB, which maintained a CPU consumption of 16 percent on low-end hardware. In the validation phase with 32 students, a significant increase in ecological knowledge was recorded, moving from 18.75 percent correct answers in the initial diagnosis to an effectiveness of up to 93 percent in tests with the tool. The usability evaluation using the SUS metric showed a satisfaction rate of 96 percent among the students. It is concluded that the integration of local computer vision and mobile gamification strengthens environmental learning at an early age, overcoming institutional connectivity limitations and providing an autonomous, scalable, and highly efficient pedagogical tool.