Hybrid Computing System for Real-Time Implementation of a Convolutional Neural Network Application

There is a drastic increase in the usage of Artificial Intelligence, Machine Learning, and Deep Learning over the past decade, and innovative applications using these technologies are being developed. The development of such applications requires a tremendous amount of data and high computational power for training and deployment. Though cloud computing is a good way for development with the increasing amount of data, cloud computing proves out to be costly and slow. Edge computing devices have become the need of the hour, which could provide high computational power locally, saving on the costs of communicating with the server and providing a faster way of processing. The issue is addressed in the presented work with the use of commercially available edge computing devices. The developed method is tested for efficiency and computation speed. For the case study, the application of concrete crack detection using a convolutional neural network is considered. Transfer learning-based approach is adopted with the use of a pre-trained inception v3 CNN model. A concrete crack dataset is prepared for training and testing the model. In order to prove the computational efficiency of edge computing, two separated models were trained ad tested. A CPU standalone model and another model compatible with CPU and edge computing device were trained and tested. The experimental results show that the CPU with edge computing device requires much less computational time as compared to the CPU standalone mode and is at least 100 times faster in computation.