Abstract:

Aim: The protection of wildlife in the Ulanba National Nature Reserve of Inner Mongolia plays a vital role in maintaining regional biodiversity. With the rapid development of artificial intelligence, deep learning has become a key tool for automating wildlife image recognition and advancing intelligent ecological monitoring. However, real-world wildlife image datasets typically exhibit a skewed distribution, where a few common species have abundant samples, while most species are underrepresented, thereby limiting the overall recognition performance of the model. 

Methods: To address this issue, this study proposes a novel method for wildlife recognition named Diff-SCC, which integrates diffusion-based data generation and feature reconstruction. Specifically, rich semantic descriptions of low-frequency categories are first generated using a large language model to guide the diffusion model in synthesizing additional samples. A multi-scale negative sample filtering strategy is then introduced to assess image quality from pixel, feature, and semantic levels, enhancing the diversity and balance of low-frequency categories’ features. Furthermore, an SCConv module is incorporated into the backbone network to improve spatial and channel modeling, focusing more effectively on foreground regions while reducing redundant computation. 

Results: This paper conducted comparative experiments on a self-built wildlife dataset from Ulanba Nature Reserve, which includes 12 wildlife categories, and on the public NACTI dataset. Experimental results show that the proposed Diff-SCC model achieves overall recognition accuracies of 78.71% and 80.84% on the two datasets, respectively. Notably, the recognition accuracy of low-ferquency classes improves by 9.96% and 9.99% over the baseline model, demonstrating the effectiveness of the proposed method in handling skewed data and recognizing rare species. 

Conclusion: The Diff-SCC model proposed in this study demonstrates strong capability in mitigating the challenges of skewed distributions in wildlife image classification. It offers a reliable and practical solution for intelligent wildlife monitoring and contributes to the advancement of biodiversity conservation.

Key words: wildlife, image classification, skewed distributions, diffusion model, feature reconstruction.