Instance Segmentation of Arowana Scales Based on Yolov8

To address the issues of low segmentation accuracy caused by complex backgrounds, low image quality, and significant scale variations in arowana scale images, we propose an instance segmentation algorithm for arowana scales called FS-YOLO based on YOLOv8. First, we replace the original C2f module in the backbone network with CRDB, which reduces the number of parameters while enhancing the feature fusion capability. Next, we introduce the HRFPN neck structure that maintains resolution to improve detail retention and strengthen the integration with semantic information. This structure enhances deep feature propagation through the AFF and BiAM attention fusion modules and completes feature fusion through the simplified redundant connection of E-FPN. Finally, an additional output head is added for segmenting small scales, increasing the model’s segmentation accuracy for the tail scales. In our created arowana scale dataset, using the CRDB backbone network instead of C2f improved mAP50 by 4.7%. Additionally, the models employing HRFPN and the added output head achieved increases of 5.5% and 3.4% in mAP50 compared to the baseline model, respectively.