Classification of Objects with Occlusions by Means of a DAG-CNN

Abstract

In the following paper, the training of a Convolutional Neural Network (CNN) type DAG (Directed Acyclic Graph) that classifies 5 tools: scalpel, pliers, scissor, screwdriver and spanner, with a 99.8% accuracy is presented, in order to perform a classification process under unexpected conditions, such as the presence of occlusions, where the visualization of activations of the convolution layers of the network allows to analyze the reasons why the network correctly or incorrectly classifies each image, for cases with different percentages of occlusion on each one of the trained categories. Subsequently, a solution for the classification of occluded objects is proposed, and a new DAG-CNN is trained, with a 99% accuracy, in order to evaluate the feasibility of the said proposal, and to observe the change of the activations against occluding percentages between 0% and 80%. Finally, the results of both networks are compared and the proposal viability and its right conditions are determined.