A method for bird individual identification based on multi-feature fusion transfer learning

Abstract

This invention provides a bird identification method based on multi-feature fusion transfer learning, belonging to the field of bird identification technology. This invention improves identification performance in complex scenes through dynamic multi-feature fusion, adaptive illumination compensation, and cross-environment transfer learning. First, dynamically adjusting the fusion ratio of texture and pose features alleviates the problem of key information loss caused by overexposure or occlusion. Second, introducing adversarial training and gradient inversion techniques eliminates the influence of illumination differences on feature distribution. Furthermore, the illumination-sensitive dynamic matching threshold design enhances the system's adaptability to extreme conditions such as strong and weak light. Combined with nonlinear reliability assessment, the reliability of the identification results is optimized, providing a highly robust and scalable identification framework for field biodiversity research.

Images