Emulação de filtros convolucionais por CNN de camada única

Abstract

This work demonstrates that a single-layer Convolutional Neural Network (CNN) can emulate classical image processing filters. The proposed approach replaces the manual and specialized tuning of these filters with an adaptive model trained directly from data. For validation, smoothing and edge detection filters were applied to datasets of pavement and vehicle license plates. Metrics such as MSE, PSNR, and SSIM confirmed that the CNN learned the filter coefficients with high fidelity, highlighting the potential for developing modular and more efficient CNNs for filtering tasks in Computer Vision.