A DEEP-FRACTURE APPROACH TO EMBEDDED MULTILAYER NEURAL NETWORKS FOR HANDWRITTEN CHARACTER RECOGNITION

Apply

A DEEP-FRACTURE APPROACH TO EMBEDDED MULTILAYER NEURAL NETWORKS FOR HANDWRITTEN CHARACTER RECOGNITION

Journal Title: Engineering and Technology Journal - Year 2022, Vol 7, Issue 07

Abstract

Low-cost microcontrollers, but also with reduced processing and memory capabilities, are highly appreciated devices for the development of specific tasks as embedded systems. The devices found in the market today provide dedicated 32-bit architectures with communication capabilities ideal for IoT applications. Some of these tasks would highly benefit from the generalization and approximation capabilities of a neural network, but current convolutional networks prove to be too complex for microcontrollers. An alternative for some cases is the use of Multilayer Neural Networks (MNNs), which, thanks to a shallower depth, reach sizes suitable for use in small embedded systems. An MNN can accommodate a complex decision surface capable of being used in high dimensionality pattern classification problems, such as handwritten character recognition. Given the lack of an explicit strategy to define the architecture of these networks in coherence with the problem to be solved, this paper conducts a performance analysis of these networks in terms of their depth, to establish criteria to determine their size in a specific application. For the evaluation, the public database MNIST has been used in conjunction with classical metrics to evaluate the performance of a classification model. The results show that the depth of the network determines to a high degree the performance of the model, and guides the appropriate selection of the architecture.

Authors and Affiliations

Fredy Martínez, Angélica Rendón

Keywords

Accuracy Confusion matrix Gradient-based learning Machine learning Multilayer neural network Optical character recognition.

EP ID EP709294
DOI 10.47191/etj/v7i7.07
Views 117
Downloads 0