FFM-ViT: an efficient fish species classification method based on deep features and transformers.

Area of Science:

• Marine biology
• Computer science
• Artificial intelligence

Background:

• Accurate fish species identification is vital for fisheries management and biodiversity conservation.
• Current classification methods struggle with small datasets and high species similarity.
• Limitations necessitate advanced computational approaches for effective fish identification.

Purpose of the Study:

• To introduce a novel deep learning model, the feature fusion module vision transformer (FFM-ViT), for enhanced fish species classification.
• To address the challenges of limited data and high similarity in existing fish identification methods.
• To improve the accuracy and efficiency of fish classification for ecological monitoring.

Main Methods:

• Developed the feature fusion module vision transformer (FFM-ViT) by integrating Mobile Inverted Bottleneck Convolution (MBConv) and Fused Mobile Inverted Bottleneck Convolution (Fuse-MBConv) blocks.
• Incorporated the channel spatial merge attention (CSMA) module to boost feature extraction and channel fusion.
• Created and utilized the Oceanfish78 dataset, comprising 78 fish categories, for model training and validation.

Main Results:

• The FFM-ViT model achieved a 90.2% accuracy rate on the Oceanfish78 dataset, significantly outperforming the standard vision transformer (ViT) model (80.4%).
• Comparative analysis on fish4knowledge and Fish31 datasets demonstrated superior performance against models like shufflenet, convnext, and swin transformer.
• Empirical results confirm the effectiveness of FFM-ViT in fish classification tasks.

Conclusions:

• The FFM-ViT model offers a robust and effective solution for fish species identification, particularly in challenging scenarios with limited data.
• The proposed method enhances high-dimensional information extraction and feature fusion, advancing deep learning applications in ichthyology.
• FFM-ViT provides valuable insights for approximate target recognition in diverse environmental contexts beyond fisheries.