TIBER-YOLO: an improved lightweight model for underwater object detection

This study presents TRACON + Inner-WIoU + BiFPN + EMPC-DetectoR (TIBER-YOLO), an improved lightweight detector for underwater object detection. Built on You Only Look Once version 8 small (YOLOv8s), the model introduces four modifications to address image degradation, small-object detection, and computational cost. First, the Triple attention Receptive-field Attention CONvolution (TRACON) module combines a triplet attention mechanism (TAM) with receptive-field attention convolution (RFAConv) to strengthen feature extraction, particularly for small targets. Second, we design inner wise intersection over union (Inner-WIoU) by integrating Inner-IoU and WIoU-v3 to improve localization accuracy and generalization. Third, a bidirectional feature pyramid network (BiFPN) improves multiscale feature fusion while reducing the number of model parameters. Finally, the efficient multiscale partial convolution detector (EMPC-Detector) combines efficient multiscale convolution (EMSConv) and partial convolution (PConv) to capture fine-grained details with lower computational complexity. TIBER-YOLO achieves mAP@ 0.5 scores of 87.1%, 86.0%, and 86.1% on the DUO, UTDAC2020, and RUOD datasets, respectively. Compared with the YOLOv8s baseline, it reduces model size, parameter count, and computational demand by 40.4%, 42.3%, and 28.2%, respectively.