Classificação de telas sensíveis ao toque de dispositivos móveis utilizando redes neurais convolucionais

Abstract

This study addresses the use of RNC for the automatic classification of touch screens from mobile devices, based on images captured in situations where it is not possible to visually identify the model or obtain direct information from the equipment. The study evaluated fifteen different models, totaling 1,500 images collected under controlled conditions. Four color spaces — RGB, HSV, LAB, and grayscale — were analyzed, and two edge detection filters (Canny and Laplace) were applied. As an experimental tool, the TensorFlow library was employed in a supervised learning process with cross-validation. The results showed that the HSV color space at a resolution of 64 × 64 pixels achieved the best performance, reaching an average accuracy of 94.4% and anaverage loss value of 0.314. The Friedman statistical test indicated significant differences among the tested methods (p < 0.05), confirming the superiority of this configuration. It is concluded that the use of RNC constitutes a feasible and efficient approach for classifying electronic components, showing potential for integration into automated systems for technical assistance and touch screen sorting.