Modelo computacional para reconhecimento de objetos em ambiente subaquático baseado em aprendizagem de máquina

Abstract

Since ancient times, man has explored the surface of the planet's waters commercial exchanges, exchanging values ​​and customs, and permeating prosperity. By expanding its borders overseas, man experienced global interconnection, where an invisible mesh of complex interrelationships connects people from all over the world. A Despite the intimacy created with ocean navigation, little is still known about the bottom of the sea. Difficulties imposed by mother nature to enter the waters combined with human limitations have contributed to keeping the underwater environment little explored. However, a unique moment for exploring the bottom of the sea begins, as several technologies in full development converge to overcome problems of access to the underwater environment, as well as the capture and processing of information from this environment. Interpretation of information captured underwater encompasses the issue of detection and recognition of objects that is the central question of this work. Machine learning, in recent years, has brought a new perspective on object recognition methods. Recently, Several machine learning and deep learning studies have been published with very relevant results. For this work, a perceptron network multi layer (MLP) was developed for detecting underwater objects from of a combination of three pre-existing datasets, resulting in a dataset with three distinct classes that include natural objects (sharks and sea turtles) and artificial objects (tubes of different sizes). The MLP simulations presented as results in a number of 888 correctly classified instances, against only 12 incorrect ones, achieving an accuracy of 98.7%. The results achieved in this work investigation shows that the proposed computational model is viable for recognizing objects in an underwater environment based on machine learning.