GAI-YOLOv8: A precision-oriented detection algorithm for identifying fish targets in sonar images

The Chinese sturgeon ( Acipenser sinensis ), a flagship species for aquatic biodiversity conservation in the Yangtze River, is currently classified as Critically Endangered due to habitat degradation and anthropogenic activities. While sonar imaging overcomes the turbidity limitations of optical methods, it faces challenges arising from complex environmental noise and the need for lightweight algorithms suitable for resource-constrained embedded devices. To address these issues, this paper introduces GAI-YOLOv8, a precision-oriented detection model. This model incorporates three key innovations: (1) the C2f-GhostDynamicConv module, which integrates lightweight Ghost architectures with dynamic convolution to adaptively enhance feature extraction while reducing parameters; (2) an ASF-P2 neck architecture that adds a high-resolution detection layer to capture fine-grained details crucial for small targets; and (3) the Inner-CIoU loss function to optimize bounding box regression and improve generalization. Evaluated on a custom dataset of 1079 sonar images, GAI-YOLOv8 achieves 80.8% precision, 80.8% recall, 84.9% [email protected], and 40.7% [email protected]:0.95, outperforming baseline YOLOv8n by 4.5%, 8.0%, 5.0%, and 3.3%, respectively. With only 1.9 M parameters, the model achieves an exceptional balance between accuracy and efficiency, surpassing other YOLO variants and high-accuracy target detect models in efficiency and accuracy trade-offs. This work advances real-time, resource-efficient underwater monitoring, offering a robust tool for ecological conservation. Future efforts will focus on hardware deployment and further architectural optimization. • Proposed GAI-YOLOv8 for precise detection of small fish in sonar imagery. • Introduced C2f-GhostDynamicConv to enhance features with fewer parameters. • Added ASF-P2 neck to improve small-target detection accuracy. • Applied Inner-CIoU loss for robust bounding box regression across scales. • Achieved 84.9% [email protected] with only 1.9 M params, enabling real-time deployment.