Post-Mortem Iris Recognition Resistant to Biological Eye Decay Processes

Mateusz Trokielewicz, Adam Czajka, Piotr MaciejewiczWACV 2020

TL;DR

This paper introduces the first end-to-end iris recognition system specifically designed for post-mortem samples, addressing the significant degradation that occurs after death. By learning custom iris filters from post-mortem data and fine-tuning a segmentation model, they achieve a substantial reduction in recognition errors compared to methods designed for live subjects.

What Problem Does This Paper Address?

Identifying deceased individuals using iris patterns is a growing area of interest, particularly for forensic applications. However, biological decay processes after death, such as tissue decomposition and changes in iris texture, severely degrade iris image quality. Existing iris recognition methods, primarily developed for live subjects, struggle with these post-mortem challenges, leading to significantly lower accuracy. This paper aims to bridge this gap by developing a robust iris recognition system tailored to the unique characteristics of post-mortem iris images.

Methods

The core of the proposed methodology lies in two key components:

  • Post-Mortem Iris-Specific Feature Representation: The authors employ a shallow Siamese convolutional neural network (CNN) to learn custom, data-driven filter kernels. These kernels are trained on pairs of iris patches from post-mortem samples. The network learns to encode iris features in a way that is more resilient to decay-induced changes. This learned filter bank is then integrated into a hybrid approach, combining it with the standard Gabor wavelets used in traditional iris recognition systems. A sequential feature selection (SFS/SBS) process is used to optimize the combination of these learned filters and existing Gabor filters, selecting the most discriminant ones.
  • Fine-tuned CNN-based Iris Segmentation: The paper also addresses the critical issue of segmentation accuracy. They fine-tune an existing CNN-based semantic segmentation model with a more diverse dataset, including ISO-compliant images, low-quality visible-light samples, and various academic iris datasets. This enhanced segmentation model provides more coherent and smoother iris masks, which is crucial for accurate feature extraction, especially when dealing with unusual deviations from standard iris patterns observed in post-mortem samples.

Datasets

  • Training Data:
    • Warsaw-BioBase-Postmortem-Iris-v1.1 and Warsaw-BioBase-Postmortem-Iris-v2 databases were used to train the custom iris feature representation filters. These datasets were processed with the new segmentation model and normalized into polar iris images.
  • The segmentation model was further fine-tuned using a combination of datasets, including:
  • Biosec baseline corpus (1200 images) * BATH database (148 images) * ND0405 database (1283 images) * UBIRIS database (1923 images) * CASIA-V4-Iris-Interval database (2639 images)
  • Evaluation Data:
    • An in-house, newly collected database of post-mortem iris images was used for evaluation. This dataset comprises images from 40 subjects, with 1094 near-infrared and 785 visible-light images, collected up to 369 hours post-demise. This dataset is subject-disjoint from the training data.

Key Insights and Ideas

  • Data-Driven Filters for Decay Resilience: The most significant insight is that learning feature extractors directly from post-mortem iris data is crucial for handling biological decay. Standard iris features are not optimized for these changes.
  • Hybrid Filter Bank Superiority: Combining learned iris-specific filters with traditional Gabor filters (via a hybrid bank) offers a better balance between robustness to decay and capturing fine-grained iris texture.
  • Segmentation is Key: The paper reinforces that accurate iris segmentation is a prerequisite for effective post-mortem iris recognition. Fine-tuning the segmentation model with diverse data significantly improves overall performance.
  • Forensic Relevance of FMR: The analysis of False Non-Match Rates (FNMR) at relaxed False Match Rate (FMR) thresholds (1% and 5%) highlights the practical utility of the system in forensic scenarios, where minimizing missed identifications is paramount.

Key Results

The proposed end-to-end system, incorporating both the improved segmentation and the new feature representation, significantly outperforms baseline methods:

  • Reduced Equal Error Rate (EER): The hybrid filter bank reduced the EER by almost a third for samples acquired up to 60 hours post-mortem, from 6.40% (original OSIRIS) to 4.39%.
  • Performance Across Time Horizons: The system consistently shows improved performance across various time horizons since death. For instance, for samples acquired up to 369 hours post-mortem, the EER was reduced from 21.36% to 19.27%.
  • Improved ROC Curves: The Receiver Operating Characteristic (ROC) curves demonstrate a clear advantage for the proposed method, especially in the low False Match Rate (FMR) regions, indicating better recognition accuracy when very few false matches are tolerated.
  • FNMR Dynamics: At a relaxed FMR of 5%, the proposed method achieves nearly perfect recognition accuracy for samples up to 24 hours post-mortem and significantly reduces errors for longer time horizons, outperforming commercial and academic benchmarks.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that neither the baseline OSIRIS nor IriCore were designed for post-mortem samples, so their lower performance is understandable. This highlights the need for specialized methods rather than a direct limitation of the proposed approach.
  • While the segmentation model is improved, the paper notes that it was trained with "manual annotations of the eye corners" for alignment, implying a degree of manual intervention in the pre-processing pipeline.
  • Future Work:
    • The authors suggest that the developed segmentation model and filter learning methodology can be directly applied to other challenging iris recognition scenarios, such as recognizing irises from diseased eyes, infants, or elderly individuals.
  • The paper provides open-source code, trained filters, and the segmentation model, facilitating further research and deployment in forensic applications.

Summary updated

BibTeX
@inproceedings{trokielewicz2020post,
  author = {Mateusz Trokielewicz and Adam Czajka and Piotr Maciejewicz},
  title = {{Post-Mortem Iris Recognition Resistant to Biological Eye Decay Processes}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2020},
}

Loading related content...

Figure