Deep-Sea Organism Detection Method Based on the SDA-HTransYOLOv8n Model

Dali Chen¹Xianpeng Shi^1*Meng Liu¹Shaojian Qiu²Zihan Zhou²

• ¹National Deep Sea Center, 738686, Qingdao, China
• ²South China Agricultural University, Guangzhou, China

Deep-sea organism detection is one of the key technologies in deep-sea resource research and conservation. However, challenges such as low recognition accuracy and insufficient robustness arise due to issues like dim lighting, severe water scattering, and blurred target features in the deep-sea environment. To address these issues, this study proposes a deep-sea organism recognition method based on an improved SDA-HTransYOLOv8n model. The model introduces significant improvements to the neck network structure of YOLOv8n. First, it replaces the traditional upsampler with an improved point sampling dynamic sampler, which adaptively adjusts the sampling rate based on the target size, reducing redundant information interference and enhancing the efficiency of image feature extraction. Second, a Semantics and Detail Infusion module (SDI) is designed to adaptively fuse feature map information across different scales, addressing the issue of small deep-sea organisms being easily overlooked while enhancing the edge and detail features of deep-sea organisms. Third, a HyperTransformer-based HT_C2f module is designed to dynamically adjust attention weights, enhancing the model's ability to capture target organism features in complex deep-sea environments and improving sensitivity to blurry and low-contrast targets. Fourth, an improved downsampling convolution module (ADown) is introduced to reduce the dimension of feature maps while retaining more key feature information, avoiding feature loss in deep-sea organism images caused by information compression during sampling. Experimental results demonstrate that, on the deep-sea organism dataset obtained by the Jiaolong manned submersible in the western Pacific Ocean, the SDA-HTransYOLOv8n model developed in this study achieves a precision of 87.6%, a mAP50 of 67.7%, and a mAP50-95 of 51.6%, respectively, representing improvements of 8.9%, 2.8%, and 1.8% compared to the original YOLOv8n model, significantly enhancing the accuracy of deep-sea organism recognition. This study effectively meets the target detection requirements in complex deep-sea environments, providing technical support for deep-sea exploration and underwater operations. Code and models are available at https://github.com/Riokuli/SDA-HTransYOLOv8n-Model.