AIN-YOLO: A lightweight YOLO network with attention-based InceptionNext and knowledge distillation for underwater object detection

Underwater object detection technology based on machine vision is the key to realizie intelligent seafood fishing. Although existing methods based on deep learning have achieved considerable advancement, such key problems as high computation complexity and low detection accuracy still exist. In order to solve these problems, a novel lightweight network named AIN-YOLO is proposed in this paper. Firstly, an attention-based InceptionNext block is designed to replace the original C2f block of YOLOv8 model by introducing a parallel branch architecture with multiple depthwise convolutions to reduce the number of parameters and the complexity of YOLOv8 model. Secondly, a nearly parameter-free improved shuffle attention module is designed to enhance the AIN-YOLO model's ability to strengthen feature channel interactions and reduce computation resource consumption. Subsequently, a knowledge distillation strategy based on bridging cross-task protocol inconsistency is introduced to train the AIN-YOLO model, facilitating effective knowledge transfer and enhancing detection accuracy. Finally, the proposed model is comprehensively analyzed and compared with six existing state-of-the-art models, and results show the proposed model exhibits better effectiveness and generalizability when detecting objects in complex underwater environment on three publicly available datasets. Moreover, model deployment experiment shows that with fewer parameters and less computation complexity, the proposed model is suitable for deploying on computation resource constrained underwater equipments.