A Lightweight Model of Underwater Object Detection Based on YOLOv8n for an Edge Computing Platform

The visual signal object detection technology of deep learning, as a high-precision perception technology, can be adopted in various image analysis applications, and it has important application prospects in the utilization and protection of marine biological resources. While the marine environment is generally far from cities where the rich computing power in cities cannot be utilized, deploying models on mobile edge devices is an efficient solution. However, because of computing resource limitations on edge devices, the workload of performing deep learning-based computationally intensive object detection on mobile edge devices is often insufficient in meeting high-precision and low-latency requirements. To address the problem of insufficient computing resources, this paper proposes a lightweight process based on a neural structure search and knowledge distillation using deep learning YOLOv8 as the baseline model. Firstly, the neural structure search algorithm was used to compress the YOLOv8 model and reduce its computational complexity. Secondly, a new knowledge distillation architecture was designed, which distills the detection head output layer and NECK feature layer to compensate for the accuracy loss caused by model reduction. When compared to YOLOv8n, the computational complexity of the lightweight model optimized in this study (in terms of floating point operations (FLOPs)) was 7.4 Gflops, which indicated a reduction of 1.3 Gflops. The multiply–accumulate operations (MACs) stood at 2.72 G, thereby illustrating a decrease of 32%; this saw an increase in the AP50, AP75, and mAP by 2.0%, 3.0%, and 1.9%, respectively. Finally, this paper designed an edge computing service architecture, and it deployed the model on the Jetson Xavier NX platform through TensorRT.