Sistema de visión artificial para el control de thrips en rosas de exportación

Abstract

This project introduces a system designed to examine pests, focusing on Thrips, in roses intended for export. These insects pose a threat to flowers, causing deformities and marks, especially on the petals, which play a crucial role in the perception of quality. Faced with this issue, it is imperative to isolate affected flowers from export batches. An advanced artificial vision system was implemented using a smartphone camera to capture images of roses. These images are transmitted to a web server and processed on a laptop through a Python program. Utilizing neural networks trained with 3200 images of red and white roses, categorized as suitable (without Thrips marks) or unsuitable, the system was validated with an additional 800 images and tested with 1284 different roses, each analyzed three times. The implementation involved the use of Keras, Tensorflow, and OpenCV libraries. Simultaneously, the physical installation of a static cabin took place, incorporating LED lighting and a mobile base to position the roses during evaluation. This design ensures controlled lighting conditions and clearly separates the software development and physical infrastructure design phases of the system. Each Convolutional Neural Network (CNN) generated from the same image database exhibited a different accuracy percentage due to the random algorithm edits, storing relevant details in .h5 files. The highest accuracy achieved in a CNN validated by images was 78%. These results, lacking a homologous system, set a significant precedent for future research. The average margin of error reached was 10.36%.