Abstract:Fine-grained image classification of wildlife species is a task of practical value and has an important role to play in the fields of endangered animal conservation, environmental protection and ecological conservation. However, the small differences between different subclasses of wildlife and the large differences within the same subclasses pose a great challenge to the classification of wildlife species. In addition, the feature extraction capability of existing methods is insufficient, ignoring the role of shallow effective features and failing to identify subtle differences between images well. To solve the above problems, this paper proposes an improved Swin Transformer architecture, called SFRSwin. Specifically, a shallow feature retention mechanism is proposed, where the mechanism consists of a branch that extracts significant features from shallow features, is used to retain important features in the shallow layers of the image, and forms a dual-stream structure with the original network. SFRSwin was trained and tested on the communal dataset Stanford Dogs and the small-scale dataset Shark species, and achieved an accuracy of 93.8% and 84.3% on the validation set, an improvement of 0.1% and 0.3% respectively over the pre-improvement period. In terms of complexity, the FLOPs only increased by 2.7% and the number of parameters only increased by 0.15%.

SFRSwin: A Shallow Significant Feature Retention Swin Transformer for Fine-Grained Image Classification of Wildlife Species.