Abstract:

With the growing demands for wildlife conservation and ecological monitoring, deep learning-based image recognition methods have been increasingly applied in wildlife research. This paper first introduces commonly used public datasets for wildlife, provides a detailed review of the applications of different deep learning techniques in wildlife image recognition, classifies these recognition methods into three levels (image-level, object-level, and pixel-level) based on task requirements, and focuses on discussing the specific implementation and technical details of the methods at each level. On this basis, the paper further explores the core challenges faced by wildlife image recognition, including various data-level issues such as uneven data quality, high annotation costs with low efficiency, and imbalanced sample distribution. Meanwhile, it also analyzes several key technical problems from the perspective of models and algorithms, including fine-grained detection, cross-domain distribution shift, class-incremental learning, zero-shot learning, and cross-modal learning. In response to the above challenges, the paper summarizes the current research progress and coping strategies, and proposes potential future development directions, aiming to provide theoretical support and methodological references for constructing an efficient, robust, and practically applicable intelligent recognition system for wildlife in real monitoring scenarios.

Key words: wildlife image recognition, deep learning, data imbalance, transfer learning, zero-shot learning, cross-modal learning