FP-Age: Face Parsing Attention Mechanism for Facial Age Estimation in the Wild

Research Background

Age estimation on facial images is a significant computer vision task with extensive applications in forensics, security, health welfare, and social media. However, due to diverse factors such as head pose, facial expressions, and occlusions, the performance of deep learning models in facial age estimation still has room for improvement. These issues are particularly pronounced in “in-the-wild” facial images under uncontrolled conditions. To enhance the robustness and accuracy of models under different conditions, the authors propose a new method aimed at incorporating facial semantic information into the age estimation process, enabling the model to effectively focus on the most informative facial regions.

Researchers and Publication Information

The primary authors of this paper include Yiming Lin, Jie Shen (corresponding author), Yujiang Wang, and Maja Pantic from Imperial College London. This paper was published in IEEE Transactions on Image Processing (volume and issue number yet to be announced, expected to be published in the future).

Research Methods

Research Process

To address the current poor performance of age estimation models in uncontrolled environments, the authors designed an FP-Age method based on a face parsing network. The core idea of this method is to improve the age estimation model by parsing facial semantic information. Specifically, the research process is divided into the following steps:

1. Face Parsing: Use a pre-trained face parsing network (such as RTNet) to extract facial semantic features.
2. Facial Parsing Attention Mechanism Module (FPA): Design a new attention mechanism module that utilizes facial semantic features for age estimation.
3. Creating the IMDB-Clean Dataset: Based on the existing IMDB-Wiki dataset, use a semi-automated method to clean the data and generate the IMDB-Clean large-scale benchmark dataset to increase the accuracy of experiments.
4. Comprehensive Experiments: Conduct extensive experiments on IMDB-Clean and other commonly used benchmark datasets and compare the performance of existing methods.

Algorithms and Methods Used

The paper adopts a method called ROI Tanh-Polar transformation to transform images, allowing better concentration on facial region features. Additionally, the paper uses convolutional neural networks (CNN) for feature extraction, combining the facial parsing network and attention mechanism to enhance model performance. The data analysis section employs label distribution learning (LDL), modeling the age estimation problem as a probability distribution problem to make the estimation results more robust and accurate.

Experimental Results

1. Creation and Validation of the IMDB-Clean Dataset:

   • The cleaned IMDB-Clean contains 287,683 images and is a challenging age estimation dataset.
   • This dataset significantly enhances model performance in uncontrolled environments.

2. Performance of FP-Age on Various Datasets:

   • On the IMDB-Clean dataset, the MAE of the FP-Age model is 4.68, with a CS5 of 63.78%, significantly better than existing state-of-the-art methods.
   • On the MORPH and CACD datasets, the performance also achieved new heights, particularly on the MORPH dataset after pre-training and fine-tuning, the MAE of FP-Age reached 1.90, setting a new record.

Conclusion and Value

This study proposes a simple yet effective method that improves the accuracy of age estimation models by incorporating facial semantic information. The results not only hold significant academic value but also have broad potential in practical applications. The introduction of the facial parsing attention mechanism (FPA) provides new ideas and references for other high-level facial analysis tasks. Meanwhile, the introduced IMDB-Clean dataset provides a new large-scale benchmark dataset for subsequent research, greatly advancing the field.

Highlights and Innovations

1. Innovative Attention Mechanism: FP-Age is the first method to achieve semantic-aware age estimation using a facial parsing attention mechanism.
2. High-Precision Age Estimation: This method achieves new optimal results on multiple benchmark datasets.
3. Data Cleaning Method: The proposed semi-automated cleaning method generates the IMDB-Clean large-scale dataset, significantly improving data quality.

Further Refinement of the Research

The authors plan to explore the issue of domain transfer between different datasets in future work. Additionally, they intend to extend the research focus to age estimation in videos, leveraging temporal information to further improve model performance.

In summary, this study brings new methods and tools to the field of facial age estimation, holding significant theoretical value and practical application prospects.

# Summary and Research Report of the "IEEE Transactions on Image Processing" Paper

## Research Background
Age estimation on facial images is a significant computer vision task with extensive applications in forensics, security, health welfare, and social media. However, due to diverse factors such as head pose, facial expressions, and occlusions, the performance of deep learning models in facial age estimation still has room for improvement. These issues are particularly pronounced in "in-the-wild" facial images under uncontrolled conditions. To enhance the robustness and accuracy of models under different conditions, the authors propose a new method aimed at incorporating facial semantic information into the age estimation process, enabling the model to effectively focus on the most informative facial regions.

## Researchers and Publication Information
The primary authors of this paper include Yiming Lin, Jie Shen (corresponding author), Yujiang Wang, and Maja Pantic from Imperial College London. This paper was published in IEEE Transactions on Image Processing (volume and issue number yet to be announced, expected to be published in the future). The DOI of the paper is 10.1109/TIP.2022.3155944.

## Research Methods
### Research Process
To address the current poor performance of age estimation models in uncontrolled environments, the authors designed an FP-Age method based on a face parsing network. The core idea of this method is to improve the age estimation model by parsing facial semantic information. Specifically, the research process is divided into the following steps:

1. **Face Parsing**: Use a pre-trained face parsing network (such as RTNet) to extract facial semantic features.
2. **Facial Parsing Attention Mechanism Module (FPA)**: Design a new attention mechanism module that utilizes facial semantic features for age estimation.
3. **Creating the IMDB-Clean Dataset**: Based on the existing IMDB-Wiki dataset, use a semi-automated method to clean the data and generate the IMDB-Clean large-scale benchmark dataset to increase the accuracy of experiments.
4. **Comprehensive Experiments**: Conduct extensive experiments on IMDB-Clean and other commonly used benchmark datasets and compare the performance of existing methods.

### Algorithms and Methods Used
The paper adopts a method called ROI Tanh-Polar transformation to transform images, allowing better concentration on facial region features. Additionally, the paper uses convolutional neural networks (CNN) for feature extraction, combining the facial parsing network and attention mechanism to enhance model performance. The data analysis section employs label distribution learning (LDL), modeling the age estimation problem as a probability distribution problem to make the estimation results more robust and accurate.

### Experimental Results
1. **Creation and Validation of the IMDB-Clean Dataset**:
    - The cleaned IMDB-Clean contains 287,683 images and is a challenging age estimation dataset.
    - This dataset significantly enhances model performance in uncontrolled environments.
    
2. **Performance of FP-Age on Various Datasets**:
    - On the IMDB-Clean dataset, the MAE of the FP-Age model is 4.68, with a CS5 of 63.78%, significantly better than existing state-of-the-art methods.
    - On the MORPH and CACD datasets, the performance also achieved new heights, particularly on the MORPH dataset after pre-training and fine-tuning, the MAE of FP-Age reached 1.90, setting a new record.

### Conclusion and Value
This study proposes a simple yet effective method that improves the accuracy of age estimation models by incorporating facial semantic information. The results not only hold significant academic value but also have broad potential in practical applications. The introduction of the facial parsing attention mechanism (FPA) provides new ideas and references for other high-level facial analysis tasks. Meanwhile, the introduced IMDB-Clean dataset provides a new large-scale benchmark dataset for subsequent research, greatly advancing the field.

### Highlights and Innovations
1. **Innovative Attention Mechanism**: FP-Age is the first method to achieve semantic-aware age estimation using a facial parsing attention mechanism.
2. **High-Precision Age Estimation**: This method achieves new optimal results on multiple benchmark datasets.
3. **Data Cleaning Method**: The proposed semi-automated cleaning method generates the IMDB-Clean large-scale dataset, significantly improving data quality.