Machine Learning in Object Recognition and Segmentation by Certain Class Neural Networks

Abstract

In this article we present a new approach to object segmentation. We use a Convolutional Neural Network with a specific architecture called U-net to recognize an object in a colored image and efficiently segment it with a new approach which we find more accurate than the one we applied in our earlier work. We focused on fast and memory-efficient segmentation process applied after training of U-net, using a single point to represent the object. The main purpose of this work is to prove that proposed simplified object representation in training of neural network leads to accurate results in meaning of counting instances of an object and in a field of segmentation process. We present results of series of experiments analyzing accuracy of computations achieved using our algorithm on nine neural networks trained on different values of various parameters.