Plankton distribution outside detection method and system based on multi-layer semantic feature diffusion modeling
By employing a multi-layer semantic feature diffusion modeling method, combined with diffusion models and deep neural networks, a latent feature space is constructed, solving the problem of identifying unknown samples in plankton image recognition systems. This enables accurate detection of samples with fine-grained distributions, improving the robustness and reliability of the system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING NORMAL UNIV AT ZHUHAI
- Filing Date
- 2026-04-28
- Publication Date
- 2026-07-17
AI Technical Summary
Existing plankton image recognition systems struggle to accurately identify unknown samples (out-of-distribution samples) in real open marine environments, especially those similar to or approximating known categories. This leads to overconfident and erroneous predictions, impacting the reliability of monitoring data and the accuracy of ecological analysis.
A multi-layer semantic feature diffusion modeling method is adopted. By combining the generative modeling capability of the diffusion model with the multi-layer semantic features of the deep neural network, a latent feature space is constructed. The multi-layer reconstruction consistency statistic is used to judge the distribution consistency of the samples, so as to achieve accurate identification of samples outside the fine-grained distribution.
It significantly improves the ability to distinguish fine-grained unknown samples, enhances the robustness and reliability of the system, achieves an AUROC of 96.47%, and reduces FPR@95TPR from 44.46% to 18.84%, demonstrating stronger stability and discriminative power under complex backgrounds and imaging perturbations.
Smart Images
Figure CN122416488A_ABST