SD-YOLOv8: An Accurate Seriola dumerili Detection Model Based on Improved YOLOv8

Area of Science:

• Computer Vision
• Aquaculture Technology
• Marine Biology

Background:

• Accurate identification of Seriola dumerili (greater amberjack) is vital for aquaculture and behavioral studies.
• Underwater environments present challenges like variable lighting and schooling fish, hindering precise species recognition.

Purpose of the Study:

• To develop an intelligent recognition model, SD-YOLOv8, for enhanced detection of Seriola dumerili.
• To improve recognition accuracy for both near and distant fish instances in complex aquatic settings.

Main Methods:

• Proposed an intelligent recognition model, SD-YOLOv8, based on the YOLOv8 network architecture.
• Incorporated a small object detection layer and head to enhance detection of smaller or distant targets.
• Utilized Deformable Convolution Network v2 (DCNv2), Bottleneck Attention Module (BAM), and redesigned Spatial Pyramid Pooling Fusion (SPPF) for improved feature extraction and fusion.
• Employed Inner-MPDIoU bounding box regression for precise localization.

Main Results:

• The SD-YOLOv8 model demonstrated significant improvements in detection accuracy and average precision.
• Accuracy increased from 89.2% to 93.2%, and average precision rose from 92.2% to 95.7%.
• The model effectively enhances recognition capabilities in challenging underwater conditions.

Conclusions:

• The developed SD-YOLOv8 model provides a reliable and accurate method for identifying Seriola dumerili.
• This advancement offers crucial technical support for Seriola dumerili aquaculture and behavioral research.
• The model's enhanced detection accuracy contributes to the broader field of fish identification technology.