Binary Classification Architecture for Edge Computing Based on Cognitive Services and Deep Neural Networks

Abstract

Systems based on computer vision and artificial intelligence are an alternative for repetitive inspection processes. However, it is possible to extend the learning capacity of these systems to classify multiple objects using edge computing. This allows combining local processing with cloud processing to expand the possibilities and reduce the response time. In this work, a classification architecture based on remote web services and local neural networks is proposed. To test this architecture, Microsoft Azure cognitive web services and its Computer Vision API have been used, combined with the use of transfer learning and ResNet 50. The cloud service allows the identification and labelling of image content, while the Edge service, based on the neural network, allows the generation of classification models for those objects not identified or incorrectly identified by the remote service. The architecture allows to extend the possibility of image recognition by integrating web services that combined with edge processing accelerate the identification process. The proposed architecture is composed of three layers; (a) a physical layer, for the mechanical and electronic structure; (b) a logical layer, which defines the interaction of the remote and local image recognition web services, and (c) an application layer, for the integration of the monitoring and control interfaces. Finally, the architecture was evaluated through functionality testing and performance metrics of classification models, as well as load and usability testing.