MDP-EEC: Model Distillation and Partitioning Enhanced End-Edge Collaboration Computing for Consumer Electronics

Deep neural network (DNN) models are widely applied in various intelligent service applications for consumer electronics, e.g., image recognition. However, the increasing scale and computational complexity of DNN models pose substantial challenges for low-latency image recognition on end-devices typically constrained by computational power. Benefiting from the low-latency of edge computing, end-edge collaboration computing can provide a promising solution. In this paper, an end-edge collaborative computing framework, named MDP-EEC, is designed for efficient image recognition, within which, knowledge distillation, model partitioning, and a minimum-latency scheduler are integrated. Specifically, an improved knowledge distillation method is proposed to compress the relatively large model into lightweight one with side branch classifiers, and the computational graph is analyzed to statically partition the lightweight model into multiple inference modules to be distributed across end-devices and edge servers. To minimize the image recognition latency within the MDP-EEC, a minimum-latency scheduler is designed to optimize the inference modules allocation. With the industrial casting product image dataset for quality inspection contained in Kaggle, experimental results have been shown to demonstrate the efficiency of investigations in the MDP-EEC.