Few-shot fine-grained fish species classification via sandwich attention CovaMNet

The task of accurately classifying marine fish species is of great importance to marine ecosystem investigations, but previously used methods were extremely labor-intensive. Computer vision approaches have the advantages of being long-term, non-destructive, non-contact and low-cost, making them ideal for this task. Due to the unique nature of the marine environment, marine fish data is difficult to collect and often of poor quality, and learning how to identify additional categories from a small sample of images is a very difficult task, meanwhile fish classification is also a fine-grained problem. Most of the existing solutions dealing with few-shot classification mainly focus on the improvement of the metric-based approaches. For few-shot classification tasks, the features extracted by CNN are sufficient for the metric-based model to make a decision, while for few-shot fine-grained classification with small inter-class differences, the CNN features might be insufficient and feature enhancement is essential. This paper proposes a novel attention network named Sandwich Attention Covariance Metric Network (SACovaMNet), which adds a new sandwich-shaped attention module to the CovaMNet based on metric learning, strengthening the CNN's ability to perform feature extraction on few-shot fine-grained fish images in a more detailed and comprehensive manner. This new model can not only capture the classification objects from the global perspective, but also extract the local subtle differences. By solving the problem of feature enhancement, this new model can accurately classify few-shot fine-grained marine fish images. Experiments demonstrate that this method outperforms state-of-the-art solutions on few-shot fine-grained fish species classification.

Automated summary

Accurately classifying marine fish species is important for marine ecosystem investigations. Previously used methods were labor-intensive, but computer vision approaches offer advantages such as long-term, non-destructive, and low-cost classification. This paper proposes a novel attention network model that enhances feature extraction from few-shot fine-grained fish images, improving classification accuracy.