Underwater-MLA: lightweight aggregated underwater object detection network based on multi-branches for embedded deployment

Abstract In response to the limited memory and computing capabilities of underwater devices, as well as the various target leakage issues caused by complex underwater environments, we propose a lightweight aggregation underwater target detection algorithm called multi-branch lightweight aggregate networks underwater (underwater-MLA). The algorithm designs a new mixed convolutional fusion with contextual attention module. It constructs a multi-branch structure that integrates mixed convolutions and contextual attention to enhance the feature extraction capability for complex underwater targets. Additionally, to address the shortcomings of information loss and significant target localization errors during detection, we propose the SPPELAN efficient aggregation module to improve the model’s perception of target localization and background discrimination ability. Furthermore, targeting the issues of blurred boundaries and sample imbalance in underwater targets, we introduce the Focaler-IoU module to focus more on the distribution of sample boundaries. Finally, we deploy the underwater-MLA algorithm on a Raspberry Pi 4B development board and utilize a camera for real-time underwater target detection. Simulation results on the URPC dataset demonstrate that this method achieves an 85.86% detection accuracy, a 2.94% improvement over the original architecture. The computational complexity and parameter count are reduced by 32.94% and 11.31%, respectively. Moreover, it achieves satisfactory target detection performance on embedded terminals. Link to open-source code: https://github.com/fanchenye/fanchenye .