A Novel Hierarchical Edge Computing Solution Based on Deep Learning for Distributed Image Recognition in IoT Systems

Traditionally, IoT systems utilize cloud computing platforms for ease of deployment. However, such platforms are resource-intensive, relatively expensive, and lead to longer response time due to network latency. Hence, cloud computing may not always be suitable for unusually large scale deployment. Various edge computing solutions have been proposed to overcome the above issues. However, most of such solutions rely upon expensive edge servers, which make them unsuitable for distributed applications like image recognition in agricultural fields. In this paper, we propose a hierarchical edge computing-based image recognition system in which the major processing is carried out at low-cost gateway devices like Raspberry Pi. As an example case, we address the issue of recognition of animals intruding in agricultural fields. We implement a dynamic learning method, in which a convolutional neural network is dynamically trained to recognize potential target classes based on a specific deployment environment. The AI detection module is then loaded on to the lowest level of edge servers on gateway devices for detection of animals and providing feedback. Experiments show that our proposed recognition system can perform offline image classification tasks with up to seven times higher accuracy and more than two times faster evaluation time in comparison with general-purpose cloud recognition systems. Besides, it consumes less than 6% of the network bandwidth and only a fraction of the energy as well as other computational resources compared with existing approaches.