Convolutional neural networks for detection of hand-written drawings
Sergio E. Valenzuela, Juan B. Calabrese, Josué Ortiz‐Medina, Claudia N. Sánchez · 2020
Convolutional Neural Networks (CNN) have been used since the late 80's. Nevertheless, until the 2000's when they begun to be popular for image classification tasks, thanks to the improvements in computation performance of electronic devices and new algorithms development. However, most of the classifiers are oriented towards the processing of real-world images. This document presents a CNN for hand-written drawings recognition. The dataset consists of 710,000 images that correspond to 71 different classes, each one with around 10,000 samples. The dataset was randomly divided into a training set (80%) and a testing set (20%). The CNN classifier achieved an accuracy of 84.79% for classifying the samples on the testing set. The classification results showed perfect identification for 10 classes, whereas 6 classes were poorly classified. It is foreseen that the results presented here can fuel applications where identification of hand-made drawings are critical, such as neuropsychological tests.