Research on jellyfish classification algorithm based on improved YOLOv8

Jellyfish are widely distributed across the world’s oceans and freshwater environments, forming an important component of marine ecosystems. Investigating the species diversity and population dynamics of jellyfish can provide valuable insights into the variations of marine organisms and their habitats. Meanwhile, the detection and recognition of jellyfish are crucial for preventing “jellyfish blooms” and maintaining ecological stability. Therefore, research on jellyfish classification algorithms holds significant ecological and societal value. We propose an improved YOLOv8-based detection model named YOLOv8-AVA for jellyfish classification and recognition. To mitigate the effects of jellyfish deformation and displacement on feature extraction, Deformable Convolutional Network version 3 (DCNv3) convolution is introduced into the Dynamic Head. To enhance the model’s focus on key jellyfish features, we introduce an Inverted Residual Mobile Block (iRMB) attention mechanism. Furthermore, we incorporate an SPP-ELAN spatial pyramid pooling structure to optimize multi-scale feature fusion, thereby enriching semantic information and enhancing computational efficiency. Additionally, we propose an Inner-Focal-MPDIoU loss function to bolster the model’s discrimination capability for challenging samples and overlapping boundaries, while also accounting for aspect ratio effects and minimizing scale sensitivity. Experimental results demonstrate that the proposed YOLOv8-AVA model achieves superior classification accuracy in jellyfish recognition tasks, indicating its practical potential and industrial value in marine ecosystem monitoring, environmental protection, and related applications.