Online Decoupled Distillation Based on Prototype Contrastive Learning for Lightweight Underwater Object Detection Models

Underwater object detection tasks face significant challenges due to the complexity of aquatic environments and underwater conditions, coupled with the limited computational resources of current underwater equipment. The characteristics of online knowledge distillation, including its ability to compress models and enhance the performance and generalization capability of lightweight models, make it a highly suitable approach for underwater object detection tasks. Despite this, the key parameter setting of online knowledge distillation is still challenging in underwater image object detection, because of the difficulties in capturing the key information that differentiates different underwater objects. Therefore, we propose an online decoupled distillation framework based on prototype contrastive learning (PCD). The core idea of PCD is to facilitate knowledge transfer among a group of networks through prototype contrastive learning and to design decoupled distillation to enable the student model to learn more comprehensive and fine-grained knowledge. In the PCD model, we employ a contrastive loss to align the distributions of the student model and the teacher model in the feature space, enhancing its semantic structural properties. This enables the student model to achieve better feature representation. Innovatively, we use the average feature vector as the prototype to execute prototype contrastive learning to ensure model stability, thereby enhancing the detection capability of lightweight models in complex environments. On detectors, we designed cross-knowledge transfer and decoupled distillation loss to make the distillation of the logit output more comprehensive and effective. Our extensive experimental results demonstrate significant improvements in underwater object detection tasks and applicability to various types of dense detectors. Our PCD can improve the average precision of ResNet-based GFL detectors by 3.0–5.3, proving the effectiveness of our PCD model.

I. Introduction

Over the past decade, convolutional neural networks (CNNs) have achieved outstanding performance across a wide range of computer vision tasks, including underwater object detection. Underwater object detection is one of the most challenging research topics in current computer vision tasks. New computer vision technology provided new insight into underwater objection detection that the development of deep learning method is available to outcome the insufficient factor in traditional methods, such as time-consuming and oversaturation. With the rise of deep learning, the development of object detection tasks has progressed rapidly, and numerous algorithms suitable for recognizing and detecting terrestrial optical images have emerged. However, the situation underwater is entirely different [1], [2].