FiberNet: um modelo de rede neural convolucional simples e eficiente

Abstract

Convolutional neural networks (CNNs) are powerful and effective tools for extracting meaningful information from images and identifying objects. However, their high computational cost can limit their adoption in scenarios with limited computational resources, such as mobile devices. To address this issue, we propose a new CNN architecture in which we include a new layer called Defiber which operates in the convolution phase of the CNN. This new layer, belonging to the down sampling strategy, aims to reduce the number of trainable parameters of the network without affecting its prediction ability. To test our approach, we created FiberNet. A prototype of a small and simple CNN that has a reduced number of trainable parameters. This resulted in a network with high inference speed and reduced computational costs. FiberNet was evaluated on two datasets, Sisal and CIFAR10. On the Sisal dataset, FiberNet achieved a precision of 96.25%. On the CIFAR10 dataset, FiberNet achieved a precision of 74.9%. Our results show that the Defiber layer is a viable alternative for building low-cost CNNs. Its application in the architecture of our model resulted in high accuracy and processing capacity, even with a reduced number of trainable parameters.