Efficient one-stage detection of shrimp larvae in complex aquaculture scenarios

The swift evolution of deep learning has greatly benefited the field of intensive aquaculture. Specifically, deep learning-based shrimp larvae detection has offered important technical assistance for counting shrimp larvae and recognizing abnormal behaviors. Firstly, the transparent bodies and small sizes of shrimp larvae, combined with complex scenarios due to variations in light intensity and water turbidity, make it challenging for current detection methods to achieve high accuracy. Secondly, deep learning-based object detection demands substantial computing power and storage space, which restricts its application on edge devices. This paper proposes an efficient one-stage shrimp larvae detection method, FAMDet, specifically designed for complex scenarios in intensive aquaculture. Firstly, different from the ordinary detection methods, it exploits an efficient FasterNet backbone, constructed with partial convolution, to extract effective multi-scale shrimp larvae features. Meanwhile, we construct an adaptively bi-directional fusion neck to integrate high-level semantic information and low-level detail information of shrimp larvae in a matter that sufficiently merges features and further mitigates noise interference. Finally, a decoupled detection head equipped with MPDIoU is used for precise bounding box regression of shrimp larvae. We collected images of shrimp larvae from multiple scenarios and labeled 108,365 targets for experiments. Compared with the ordinary detection methods (Faster RCNN, SSD, RetinaNet, CenterNet, FCOS, DETR, and YOLOX_s), FAMDet has obtained considerable advantages in accuracy, speed, and complexity. Compared with the outstanding one-stage method YOLOv8s, it has improved accuracy while reducing 57 % parameters, 37 % FLOPs, 22 % inference latency per image on CPU, and 56 % storage overhead. Furthermore, FAMDet has still outperformed multiple lightweight methods (EfficientDet, RT-DETR, GhostNetV2, EfficientFormerV2, EfficientViT, and MobileNetV4). In addition, we conducted experiments on the public dataset (VOC 07 + 12) to further verify the effectiveness of FAMDet. Consequently, the proposed method can effectively alleviate the limitations faced by resource-constrained devices and achieve superior shrimp larvae detection results.