Lightweight Reasoning Acceleration Algorithm of Deep Neural Network for Edge Computing Environment

In this paper, a lightweight reasoning acceleration algorithm based on deep neural network (DNN) for edge computing environment is studied. Edge intelligent devices have certain computing power and can provide low latency and high-efficiency intelligent services on the terminal side close to the data source. However, to improve model accuracy, modern DNNs typically adopt deeper and more complex network structures, resulting in a significant increase in parameter count, storage overhead, and computational load. To effectively reduce terminal energy consumption while meeting user real-time requirements, this paper proposes a DNN inference optimization scheduling method based on edge collaboration. This method divides the DNN network layer into fine-grained layers and combines the collaborative computing capabilities of edge servers and terminal devices to search for the optimal inter layer allocation strategy, thereby minimizing overall inference latency. Furthermore, this article proposes a task offloading mechanism that faces multiple resource constraints and integrates an improved particle swarm optimization algorithm (PSO) for scheduling and solving. While meeting the delay constraints, it significantly reduces the energy consumption of terminal devices. The results indicate that the proposed method can effectively shorten inference time and reduce terminal energy consumption.