Giant Panda Face Recognition Using Small Dataset

Wojciech Michal Matkowski, Adams Wai-Kin Kong, Han Su, Peng Chen, Rong Hou, Zhihe Zhang

Abstract

Giant panda (panda) is a highly endangered animal. Significant efforts and resources have been put on panda conservation. To measure effectiveness of conservation schemes, estimating its population size in wild is an important task. The current population estimation approaches, including capture-recapture, human visual identification and collection of DNA from hair or feces, are invasive, subjective, costly or even dangerous to the workers who perform these tasks in wild. Cameras have been widely installed in the regions where pandas live. It opens a new possibility for non-invasive image based panda recognition. Panda face recognition is naturally a small dataset problem, because of the number of pandas in the world and the number of qualified images captured by the cameras in each encounter. In this paper, a panda face recognition algorithm, which includes alignment, large feature set extraction and matching is proposed and evaluated on a dataset consisting of 163 images. The experimental results are encouraging.

Introduction

Conservationists require reliable methods to track individual giant pandas to estimate population sizes and study behavior. Traditional monitoring techniques often involve invasive physical tagging, subjective human visual identification, or expensive and potentially dangerous DNA collection from scat in remote, rugged terrain.

Distinctive patterns in zebras and tigers contrast with the subtle facial differences in giant pandas.

While camera traps offer a non-invasive alternative, identifying individual pandas from images is exceptionally difficult. Unlike species such as zebras or tigers that possess distinct, "fingerprint-like" stripe patterns, pandas appear remarkably similar to the human eye. Furthermore, computer vision for this task faces a "small dataset" hurdle. With approximately 1,864 pandas remaining in the wild and limited high-quality images per encounter, standard deep learning architectures—which typically require millions of training images—often overfit and fail to generalize.

Methodology

The proposed recognition pipeline is specifically designed to maximize information extraction from limited data through a multi-stage process:

1. Image Alignment

To ensure features are compared consistently, probe images are aligned to gallery images using Coherent Point Drift (CPD). A Sobel edge detector identifies key points, and the CPD algorithm calculates an affine transformation to warp images into a common coordinate space.

2. Multi-Scale Feature Extraction

Rather than relying on a single global view, the system extracts Local Binary Patterns (LBP) and Gabor histograms across seven different grid configurations. These grids range in density from $2 \times 2$ to $7 \times 5$ blocks, allowing the model to capture both fine textures and broader spatial structures. This results in a high-dimensional feature vector of 15,374 dimensions per image.

3. Classification via PLS Regression

To manage high dimensionality and class imbalance, Partial Least Square (PLS) regression is employed. A "one-against-all" approach builds robust classifiers for each individual panda, mitigating the overfitting common in deep learning models when data is scarce.

Dataset

The study utilized a dataset from the Chengdu Research Base of Giant Panda Breeding, containing 163 images of 28 individual pandas. These images feature significant variations in pose, lighting, and occlusions (e.g., pandas eating bamboo). Frontal face regions were manually cropped and normalized for the study.

Results

The algorithm was evaluated using leave-one-out cross-validation and compared against five state-of-the-art deep learning architectures: AlexNet, GoogLeNet, ResNet-50, VGG-16, and VGG-Face.

  • Identification: The proposed method achieved a 71.17% Rank-1 accuracy, outperforming the second-best model (ResNet-50 at 62.58%) by 8.59%.
  • Verification: At a 1% False Acceptance Rate (FAR), the system achieved a 65.64% True Acceptance Rate (TAR), roughly 6.4% higher than ResNet-50.
  • Efficiency: The system proved robust against the overfitting that plagued deep learning models, which were only fine-tuned on their last layers due to the small sample size.

Conclusion and Future Work

This research establishes a viable framework for non-invasive panda identification. While successful, the current reliance on manual cropping is a limitation. Future work will focus on integrating automated panda face detection and refining features to include specific facial components like ear shapes and eye-patch nuances. As more data is collected, the researchers aim to further scale the system for broader field application.

Summary updated

Abstract

Giant panda (panda) is a highly endangered animal. Significant efforts and resources have been put on panda conservation. To measure effectiveness of conservation schemes, estimating its population size in wild is an important task. The current population estimation approaches, including capture-recapture, human visual identification and collection of DNA from hair or feces, are invasive, subjective, costly or even dangerous to the workers who perform these tasks in wild. Cameras have been widely installed in the regions where pandas live. It opens a new possibility for non-invasive image based panda recognition. Panda face recognition is naturally a small dataset problem, because of the number of pandas in the world and the number of qualified images captured by the cameras in each encounter. In this paper, a panda face recognition algorithm, which includes alignment, large feature set extraction and matching is proposed and evaluated on a dataset consisting of 163 images. The experimental results are encouraging.
BibTeX
@inproceedings{matkowski2019giant,
  author = {Wojciech Michal Matkowski and Adams Wai-Kin Kong and Han Su and Peng Chen and Rong Hou and Zhihe Zhang},
  title = {{Giant Panda Face Recognition Using Small Dataset}},
  booktitle = {ICIP},
  year = {2019},
  url = {https://openreview.net/forum?id=GTQkwH3HegQ},
}
Figure