RSF-YOLOv8: a re-parameterized multi-branch feature augmentation network for polarization-guided underwater object detection

Underwater object detection techniques for marine resource exploration, underwater engineering inspection, and AUV intelligent operations face three major demands: high precision, low latency, and embeddable deployment. In natural underwater environments, complex interferences such as water scattering, wavelength-selective attenuation, illumination distortion, and target occlusion commonly exist. As a result, traditional optical imaging methods often suffer from inherent drawbacks, including image blurring, low contrast, and edge degradation. Polarization imaging can separate scattered stray light and preserve the intrinsic polarization texture of targets. Therefore, it exhibits significant advantages in highly turbid and low-light underwater scenes. However, existing lightweight detection models are not structurally adapted to the unique features of polarization images, leading to weak multi-scale modeling capacity, suppression of weak target features, and high miss rates for small targets. To address these issues, a novel underwater object detection model termed RSF-YOLOv8 is proposed based on the YOLOv8n baseline. Three core innovative modules are designed to optimize the model: 1) An EfficientRep re-parameterized multi-branch convolution module is constructed. By combining multi-scale parallel convolutions with structural reparameterization, the multi-scale feature extraction capability is enhanced without increasing inference cost, thereby adapting to the large-scale span and irregular morphology of underwater targets. 2) An EfficientSE dual-pooling channel attention module is designed. Global average pooling and max pooling are fused to replace the traditional single-pooling SE structure so that target feature channels are adaptively enhanced while redundant background channels are suppressed, improving the recognition capability for low-contrast polarized targets. 3) A FASFF4 (Four-scale Adaptive Spatial Feature Fusion) module is established. Unlike the original PAN-FPN which only performs simple lateral connections across three scales(1/32, 1/16, 1/8) and lacks fine-grained 1/4 detail preservation, our FASFF4 introduces a dedicated 1/4-scale branch and a bidirectional cross-scale interaction pathway that explicitly transmits high-level semantics downward while compensating detailed edge cues upward through an additional bottom-up path. This four-scale bidirectional design alleviates feature dilution and localization offset for very small underwater target, which are inherently prone to vanishing in conventional three-scale fusion. Systematic experiments are conducted on a self-constructed polarized underwater organism dataset(PUOD) and four public/self-constructed generalization datasets(UDD, DUO, URPC2020, BMPD). The results show that, compared with the baseline YOLOv8n, the proposed RSF-YOLOv8 achieves a precision of 0.9394(from 0.9299), an mAP@0.5 of 0.9939(from 0.9839), and a mAP@[0.5:0.95] of 0.7125(from 0.6792) on the PUOD dataset. Although the 1.02% gains in precision and mAP@0.5 appear marginal in absolute terms, they represent a relative reduction of 13.4% and 6.2% in false positives and misclassifications, respectively, given that the baseline already operates in a high-accuracy regime(>0.98). More critically, the 4.90% improvement in mAP@[0.5:0.95] reflects a substantial enhancement in strict localization accuracy, which directly translates to fewer missed small targets in real AUV operations. Ablation studies validate the effectiveness and synergistic effects of each module, while generalization experiments demonstrate strong robustness across different underwater scenarios. This research provides a high-precision and real-time detection solution for AUVs equipped with polarization cameras, achieving an inference speed of 62.3FPS and 16.0ms single-image latency on RTX4060 GPU, as well as 18.2FPS on Intel i9-13900HX CPU after structural reparameterization. With only 3.1M parameters and 8.1 GFLOPs at inference, the model provides a computationally efficient architectural foundation that is well-suited for resource-constrained underwater platforms.