Porting deep neural networks on the edge via dynamic K-means compression: A case study of plant disease detection

Cyber Physical Systems (CPS) totally revolutionized the way we interact with the world providing useful services that can support the human being in many aspects of his life. Artificial Intelligence (AI) is another important player for bringing intelligence to CPS and allows the realization of Intelligent Cyber Physical Systems where smart applications can run. However, the constrained hardware of these devices in terms of memory and computing power makes challenging the deployment and execution of powerful algorithms (e.g., deep neural networks). To address this problem, modern solutions involve the use of compression techniques to reduce the memory footprint of deep learning models while saving the accuracy performance. The proposed work focuses on plant disease detection which represents one of the biggest challenges in smart agriculture; in such a context, the possibility to perform a timely diagnosis on crops suspected to be infected can avoid the spread of diseases, thus saving a lot of time and money during the plantation works. In this paper, we realized an intelligent CPS on top of which we implemented an AI application, called Deep Leaf that exploits Convolutional Neural Networks to detect the main biotic stresses affecting crops. To meet the hardware requirements of the Edge device running our application, we propose a novel dynamic compression algorithm based on K-Means for the reduction of models footprint. Experimental results show that our detector is able to correctly classify the plant health condition with an accuracy of 95% and demonstrate the effectiveness of the proposed compression algorithm which is able to maintain the same accuracy of the original 32 bit float model, with an overall memory size reduction of about 85.2%.