Domain generalization for sea cucumber detection: Tackling background color variability in aquaculture settings

In underwater aquaculture environments, variations in image styles caused by factors such as lighting conditions, water quality, and plankton presence introduce significant domain shift challenges for object detection tasks. To address these challenges, this paper proposes a novel sea cucumber detection model, UICTDG-YOLO, which utilizes color domain generalization techniques. Specifically, we employ a frequency-domain enhancement method based on the Fourier transform. This technique reconstructs images by perturbing the amplitude spectrum while retaining the original phase spectrum, effectively enhancing color consistency across diverse aquatic environments. Additionally, a parameterized compensation mechanism is integrated to preserve target information, thereby augmenting the dataset, increasing domain diversity, and improving the model’s generalization capability. One of the major challenges in sea cucumber detection is distinguishing target features from complex background elements. To address this, we integrate a SENetv2-based compression and aggregation network into the model backbone, enhancing its ability to extract key target features from cluttered underwater environments. Furthermore, considering the substantial shape and scale variations of sea cucumbers across different aquaculture environments, as well as the presence of background objects with textures resembling sea cucumbers, we incorporate a Bidirectional Feature Pyramid Network (BiFPN) module into the network’s neck. This module facilitates multi-scale feature fusion, improving detection accuracy across varying scales and effectively reducing background interference. Given the class imbalance in the dataset, we employ the Focal-GIoU loss function to address the imbalance between positive and negative samples while improving the accuracy of bounding box regression. Experimental results demonstrate that UICTDG-YOLO significantly outperforms the baseline model, achieving a 5.8% improvement in mean average precision (mAP), a 5.8% improvement in precision (P), and a 6.2% improvement in recall (R). The model consists of 10.4 million parameters, with a computational load of 27.5 GFLOPs and a detection rate of 28.5 frames per second. When compared to prominent object detection models, including Faster R-CNN, YOLO series models, WQTDG-YOLO, and OA-DG, UICTDG-YOLO demonstrates clear advantages in sea cucumber detection tasks within real aquaculture environments. This model provides valuable technical insights and practical applications for the scientific farming of sea cucumbers.