ViT-CatNet: An End-to-End Vision Transformer Architecture for Instance-Level Feline Identification

In recent years, the recognition of individual animals has been used more and more. However, although TNR plays an important role in the management of stray cats in supermarkets, the effect of this method lacks instance-level monitoring on population of cats. We present VIT-CatNet, an end-to-end vision Transformer architecture integrated with dynamic loss selection that applies VIT technology to detect individual stray cats via cameras. By leveraging the self-attention mechanism, VITCatNet captures global dependencies in the image. A multi-source dataset, including public repositories, online videos, and community-collected images, ensures a realistic simulation of a stray cat observation scenario. This dataset was used to train the model. Experiments show that Focal Loss achieves superior performance, balanced stability, and robustness while solving the low-sample problem. This work provides scalable, resource efficient solutions to urban and ecological challenges.