Criss-cross global interaction-based selective attention in YOLO for underwater object detection

Shen, Xipeng; Wang, Huibing; Li, Yafeng; Gao, Tianzhu; Fu, Xianping

doi:10.1007/s11042-023-16311-y

Cited by 4 publications

(1 citation statement)

References 40 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Attention mechanisms stem from the study of human vision and its ability to sift through large amounts of information to find important data. Shen et al [17] proposed the crisscross global interaction strategy (CGIS) for versions of YOLO detectors in order to minimize the interference of the underwater background with the detected target. Yu et al [18] proposed YOLOv7-net, which added bi-level routing attention (BRA) and a new coordinated attention module (RFCAConv) to YOLOv7.…”

Section: Introductionmentioning

confidence: 99%

Underwater Object Detection in Marine Ranching Based on Improved YOLOv8

Jia,

Lv,

Chen

et al. 2023

JMSE

View full text Add to dashboard Cite

The aquaculture of marine ranching is of great significance for scientific aquaculture and the practice of statistically grasping existing information on the types of living marine resources and their density. However, underwater environments are complex, and there are many small and overlapping targets for marine organisms, which seriously affects the performance of detectors. To overcome these issues, we attempted to improve the YOLOv8 detector. The InceptionNeXt block was used in the backbone to enhance the feature extraction capabilities of the network. Subsequently, a separate and enhanced attention module (SEAM) was added to the neck to enhance the detection of overlapping targets. Moreover, the normalized Wasserstein distance (NWD) loss was proportionally added to the original CIoU loss to improve the detection of small targets. Data augmentation methods were used to improve the dataset during training to enhance the robustness of the network. The experimental results showed that the improved YOLOv8 achieved the mAP of 84.5%, which was an improvement over the original YOLOv8 of approximately 6.2%. Meanwhile, there were no significant increases in the numbers of parameters and computations. This detector can be applied on platforms for seafloor observation experiments in the field of marine ranching to complete the task of real-time detection of marine organisms.

show abstract

Section: Introductionmentioning

confidence: 99%