Author(s): Xing Xuanwei; Wang Congcong; Xue Yuan; Zhang Yongxian; Li Xinyang; Xu Mengzhen

Linked Author(s):

Keywords: Golden mussel (Limnoperna fortunei); Target detection; Aquatic ecology; Lightweight model

Abstract: The purpose of this study is to solve the problem of target detection of gold mussel larvae, and to establish a detection model of gold mussel larvae recognition with low computational complexity and high detection performance. The biofouling caused by golden mussels (Limnoperna fortunei) poses significant threats to the safety of water diversion projects and economic production. During their planktonic larval stage, golden mussels can be transported in large quantities through water flows into conveyance facilities, where they establish, reproduce, and grow, making eradication exceptionally difficult once colonization occurs. Effective control strategies must target the most critical larval stage. Rapid and accurate identification of the growth stage of golden mussel larvae, followed by the direct implementation of targeted control measures based on the determined results, plays a positive role in improving efficiency, preventing further spread and attachment of golden mussels, and thereby reducing engineering losses. To address the challenges of manual identification and detection of golden mussel larvae, this study introduces Mussel-ID, an innovative, efficient, and lightweight deep learning model specifically designed for detecting golden mussel larvae. Built upon the YOLO v11 architecture, Mussel-ID integrates advanced techniques to optimize performance while minimizing computational complexity, enabling deployment on resource-constrained devices. The model incorporates lightweight modules such as DSC Conv and DRSRM for reduced size and enhanced processing speed, attention mechanisms like ECA and Element-wise Multiplication to refine feature focus and improve detection accuracy, and an affine coordinate transformation module to precisely measure larval dimensions for further analysis and control applications. Extensive experiments validate the exceptional performance of Mussel-ID, achieving a high mAP@0.5 of 98.5% and mAP@0.5: 0.95 of 91.3%. Its compact design and low computational demands make it suitable for real-time applications on edge devices. This study demonstrates the potential of Mussel-ID to revolutionize golden mussel detection, offering a robust solution for effective biofouling control and contributing to sustainable environmental management.

DOI: https://doi.org/10.64697/978-90-835589-7-4_41WC-P1780-cd

Year: 2025