LGCN: Learnable Gabor Convolution Network for Human Gender Recognition in the Wild

Peng Chen, Weijun Li, Linjun Sun, Xin Ning, Lina Yu, Liping ZhangIEICE Trans. Information and Systems 2019 · OpenReview

Abstract

Human gender recognition in the wild is a challenging task due to complex face variations, such as poses, lighting, occlusions, etc. In this letter, learnable Gabor convolutional network (LGCN), a new neural network computing framework for gender recognition was proposed. In LGCN, a learnable Gabor filter (LGF) is introduced and combined with the convolutional neural network (CNN). Specifically, the proposed framework is constructed by replacing some first layer convolutional kernels of a standard CNN with LGFs. Here, LGFs learn intrinsic parameters by using standard back propagation method, so that the values of those parameters are no longer fixed by experience as traditional methods, but can be modified by self-learning automatically. In addition, the performance of LGCN in gender recognition is further improved by applying a proposed feature combination strategy. The experimental results demonstrate that, compared to the standard CNNs with identical network architecture, our approach achieves better performance on three challenging public datasets without introducing any sacrifice in parameter size.

Introduction

Identifying gender from facial images "in the wild"—where lighting, poses, and occlusions vary significantly—remains a difficult computer vision task. Historically, researchers have relied on either hand-crafted features like Gabor filters, which are robust but require tedious manual tuning, or deep learning (CNNs), which learn features automatically but can overfit on limited data.

Previous attempts to merge these approaches often resulted in larger, slower models. The Learnable Gabor Convolution Network (LGCN) addresses this by embedding the structural "prior knowledge" of Gabor filters directly into the CNN architecture while allowing the network to tune the filters' mathematical parameters automatically via backpropagation.

Methods

The core innovation is the Learnable Gabor Filter (LGF). Standard 2D Gabor filters are defined by parameters like wavelength ($\lambda$), orientation ($\theta$), and phase offset ($\psi$). By deriving partial derivatives for these parameters, the authors made them compatible with standard backpropagation.

LGCN Architecture

The LGCN uses an AlexNet-like structure but modifies the first layer. Instead of standard random convolutional kernels, it employs 96 LGFs to capture diverse frequency and orientation responses from color images.

Feature Combination Strategy (LGCN-C)

Recognizing that Gabor filters excel at textures but might miss other patterns, the authors proposed a "combined" version, LGCN-C. In this architecture, the first layer is a mix: some kernels are LGFs, while others are standard convolutional kernels. This allows the network to capture both structured textures and arbitrary, unknown patterns simultaneously.

Architecture of LGCN and LGCN-C models showing the integration of Learnable Gabor Filters.

Experimental Results

The LGCN and LGCN-C models were evaluated on three prominent benchmarks: Adience, CelebA, and LFW. The results consistently showed that these models outperformed both standard CNNs and previous state-of-the-art methods.

Performance Benchmarks

As shown in the table below, LGCN-C achieved superior accuracy across all datasets. Notably, on the Adience dataset, it reached 88.8% accuracy, surpassing the previous best of 86.8%. On CelebA, it achieved 98.52%, outperforming other high-performing models like PaW and AFFACT.

Comparison of LGCN and LGCN-C with state-of-the-art methods across three major datasets.

Efficiency and Stability

Beyond accuracy, LGCN offers significant efficiency. A standard 5x5 convolutional kernel has 25 learnable parameters, whereas a Gabor filter is defined by only 5. Replacing standard kernels with LGFs reduced the parameter count of those specific kernels by 80%. Furthermore, the self-learning capability of LGFs ensures the model converges on optimal parameter values, providing more stable performance than manual trial-and-error tuning.

Conclusion

LGCN successfully combines the texture-capturing strengths of bio-inspired Gabor filters with the flexibility of deep learning. By making Gabor parameters learnable, the framework achieves state-of-the-art performance in unconstrained gender recognition tasks without increasing model size. Future work will focus on the joint learning of all multi-parameters of the Gabor filters to further specialize feature extraction for various facial attribute tasks.

Summary updated

Abstract

Human gender recognition in the wild is a challenging task due to complex face variations, such as poses, lighting, occlusions, etc. In this letter, learnable Gabor convolutional network (LGCN), a new neural network computing framework for gender recognition was proposed. In LGCN, a learnable Gabor filter (LGF) is introduced and combined with the convolutional neural network (CNN). Specifically, the proposed framework is constructed by replacing some first layer convolutional kernels of a standard CNN with LGFs. Here, LGFs learn intrinsic parameters by using standard back propagation method, so that the values of those parameters are no longer fixed by experience as traditional methods, but can be modified by self-learning automatically. In addition, the performance of LGCN in gender recognition is further improved by applying a proposed feature combination strategy. The experimental results demonstrate that, compared to the standard CNNs with identical network architecture, our approach achieves better performance on three challenging public datasets without introducing any sacrifice in parameter size.
BibTeX
@article{chen2019lgcn,
  author = {Peng Chen and Weijun Li and Linjun Sun and Xin Ning and Lina Yu and Liping Zhang},
  title = {{LGCN: Learnable Gabor Convolution Network for Human Gender Recognition in the Wild}},
  journal = {IEICE Trans. Information and Systems},
  year = {2019},
  url = {https://openreview.net/forum?id=r75DJyBePj},
}
Figure