Lightweight Infrared Maritime Target Detection Algorithm Based on Fusion Shuffling Convolutional Attention

Infrared maritime target detection is fundamental in maritime rescue and maritime security fields. We collected an infrared maritime target dataset and found that the texture details of infrared images are blurry, leading to poor detection performance by algorithms. To address this, we propose a lightweight infrared maritime target detection algorithm based on fusion shuffling convolutional attention. This algorithm employs the lightweight network EfficientformerV2 as the backbone for feature extraction, reducing model parameters while ensuring feature extraction capability. Additionally, the algorithm introduces the Dysample upsampler to efficiently restore the resolution of infrared image features with minimal computational cost, mitigating the accuracy loss caused by lightweighting. Finally, we propose the lightweight mixed attention mechanism called fusion shuffling convolutional attention, which effectively extracts spatial and channel features of images with only a slight increase in parameters. Experiments were conducted using our self-collected infrared maritime target dataset to train the model. Results show that, compared to RT-DETR, our algorithm's [email protected] increased by 1.2%, reaching 87.5%, while the number of parameters and computational cost decreased by 63% and 71.4%, respectively. The improved algorithm enhances detection performance while significantly reducing parameters and computational cost.