ChildrEN SafEty and Rescue (CENSER) System for Trafficked Children from Brothels in India

Raghu Vamshi Hemadri, Amarjot Singh, Ajeet SinghAAAI 2022 · DOI

Introduction

Child trafficking is a global crisis of immense proportions, with millions of children forced into labor or exploitation. In India, NGOs such as Guria work tirelessly to rescue these children, but they face a significant technical hurdle: identification. When a child is trafficked, years often pass before they are located. During this time, their facial features change significantly due to aging. Furthermore, the only available records are often outdated photographs, hand-drawn sketches from memory, or photos of biological relatives.

Standard face recognition systems frequently fail under these "unconstrained" conditions, which include poor lighting, low-quality hidden camera footage, and extreme age gaps. The CENSER (ChildrEN SafEty and Rescue) system was developed to bridge this gap, providing a robust, real-time identification tool designed for high-stakes rescue operations.

The CENSER Framework

The core of the CENSER system is the Memory Augmented ScatterNet ResNet Hybrid (MSRHN). This architecture is a sophisticated blend of traditional signal processing and modern deep learning designed to work effectively even with limited data.

  • ScatterNet (Front-end): This layer utilizes a handcrafted "Dual-Tree Complex Wavelet Transform" (DTCWT). Unlike standard Convolutional Neural Network (CNN) layers that must learn filters from scratch, ScatterNet uses predefined wavelets to extract invariant edge features. This allows the model to learn effectively from the relatively small datasets typical of trafficking cases.
  • Inception ResNet (Back-end): High-level features extracted by the ScatterNet are processed by an Inception ResNet architecture to learn complex semantic representations.
  • Memory Buffer: To handle "one-shot" learning, the system includes a memory buffer. If a new face is encountered frequently but is not in the training set, its features are stored here, allowing the system to recognize the individual without requiring a full retraining cycle.
Original vs. artificially aged sketches for identification.

To ensure accuracy in rescue missions, the authors implemented three critical safeguards: gender determination to filter out incorrect matches, temporal consistency (requiring a face to appear in three consecutive video frames), and uncertainty estimation via "dropout at test time" to discard low-confidence matches.

Data and Kinship Verification

The researchers introduced the CENSER Dataset, which contains images of 358 subjects categorized by age progression, kinship, and sketches. This dataset is uniquely challenging as it includes real-world artifacts like water damage, blur, and poor illumination.

One of the system's most innovative features is its use of kinship verification. If a child was abducted at such a young age that no photos exist, the system can match the child’s face to the features of their parents or siblings to establish identity. This serves as a vital fallback when direct identification is impossible.

CENSER dataset examples showing kinship and age progression.

Results and Real-World Impact

The MSRHN model achieved a remarkable 93.175% accuracy on the CENSER dataset, significantly outperforming previous state-of-the-art methods such as CARC (83.47%) and LF-CNNs (88.14%). Beyond accuracy, the system is highly efficient, processing video at 12–18 frames per second (fps) on standard cloud GPUs. This efficiency makes it viable for real-time use by field volunteers using a mobile app connected to an AWS backend via Apache Kafka and Spark.

Most importantly, the system has demonstrated tangible real-world success. At the time of publication, the CENSER system had already played a pivotal role in the rescue of 13 children from exploitative environments.

Conclusion and Future Work

While the CENSER system represents a major leap forward, challenges remain in extreme unconstrained environments, such as near-total darkness or extreme camera angles. Future research will focus on improving robustness in these edge cases and refining the "Memory Buffer" logic to better manage the transition from temporary feature storage to permanent model integration, all while maintaining strict privacy standards for the children involved.

Summary updated

BibTeX
@inproceedings{hemadri2022children,
  author = {Raghu Vamshi Hemadri and Amarjot Singh and Ajeet Singh},
  title = {{ChildrEN SafEty and Rescue (CENSER) System for Trafficked Children from Brothels in India}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i11.21449},
}
Figure