Abstract
Maritime mobile edge computing (MMEC) technology facilitates the deployment of computationally intensive object detection tasks on Maritime Internet of Things (MIoT) devices with limited computing resources. However, the dynamic marine network and environmental interference in feature extraction adversely affect detection accuracy and cause delay. In this paper, we propose a cumulative confidence-driven joint scheduling strategy for image detection tasks in MMEC scenarios. The strategy employs lightweight and full models as the detection framework. Through an adaptive decision-making scheme for marine device image recognition, the proposed strategy accumulates results from different models within the framework to ensure quality of service (QoS). To obtain a dynamic offloading strategy that minimizes the total system cost of latency and energy consumption, the problem is divided into two subproblems, and a chemical reaction optimization algorithm is used to reduce the computational complexity. Then, a state normalization action project deep deterministic policy gradient (SNAP-DDPG) algorithm is proposed to handle environmental dynamics, achieving minimized system cost with satisfied QoS. The simulation results indicate that, compared to existing algorithms, the proposed SNAP-DDPG algorithm keeps object detection confidence, with latency reduced by 34.78%.