Classification of Scenes Using Specially Designed Convolutional Neural Networks for Detecting Robotic Environments.

This article addresses the task of classifying scenes in typical indoor environments navigated by robots, using pre-trained convolutional neural network - CNN models with ImageNET and PLACES 365 datasets. The implemented models are the CNN VGG16 and RESNET50 architectures, which underwent various manipulations such as freezing and modification of intermediate layers. The performance of these networks in scene recognition tasks was analyzed, where a “scene” refers to an image that captures a partial or complete view of an indoor environment, displaying objects and their spatial relationships, and is similar to the scenes that mobile robots must perceive in their navigation task. To achieve this objective, a set of custom images JUNIO20V1, FEBR20V3, JUNIO20V3 and others, as well as 2 state-of-the-art image sets 15-Scenes and Sports, were subjected to various manipulations such as rotations, perspectives, sectioning, and lighting changes, creating a diverse image bank for both training and testing, allowing the analysis of the performance of the different modified CNN structures. The obtained results demonstrate the strengths of the different CNN structures in scene recognition tasks for application in indoor environments where mobile robots move.