Examinando por Autor "Garcia-Espinosa, Francisco J."

Mostrando 1 - 2 de 2

Logic Neural Networks for Efficient FPGA Implementation
(Institute of Electrical and Electronics Engineers, 2024-11-07) Ramírez, Iván; Garcia-Espinosa, Francisco J.; Concha, David; Aranda, Luis Alberto
Logic Neural Networks (LNNs) represent a new paradigm for implementing neural networks in hardware devices such as Field-Programmable Gate Arrays (FPGAs). These network architectures exhibit unique attributes that can leverage the inherent parallelism of FPGAs, enabling the development of networks characterized by low power consumption and fast inference capabilities. Despite their potential advantages, the relative novelty of LNNs poses a challenge, as there are currently no established guidelines for defining their architectures. In this paper, we present a comprehensive study of LNNs, aiming to address the existing gap in understanding and guide decision-making during the design phase. Through systematic experimentation and analysis, we explore various aspects of logic networks, including their impact on inference time, power consumption, and overall simplicity. The findings derived from these experiments provide valuable insights for the creation of improved networks, thereby paving the way for further advancements in this field
Visual classification of dumpsters with capsule networks
(ACS, 2022) Garcia-Espinosa, Francisco J.; Concha, David; Pantrigo, Juan J.; Cuesta-Infante, Alfredo
Garbage management is an essential task in the everyday life of a city. In many countries, dumpsters are owned and deployed by the public administration. An updated what-and-where list is in the core of the decision making process when it comes to remove or renew them. Moreover, it may give extra information to other analytics in a smart city context. In this paper, we present a capsule network-based architecture to automate the visual classification of dumpsters. We propose different network hyperparameter settings, such as reducing convolutional kernel size and increasing convolution layers. We also try several data augmentation strategies, as crop and flip image transformations. We succeed in reducing the number of network parameters by 85% with respect to the best previous method, thus decreasing the required training time and making the whole process suitable for low cost and embedded software architectures. In addition, the paper provides an extensive experimental analysis including an ablation study that illustrates the contribution of each component in the proposed method. Our proposal is compared with the state-of-the-art method, which is based on a Google Inception V3 architecture pretrained with Imagenet. Experimental results show that our proposal achieves a 95.35% accuracy, 2.35% over the previous best method.