[Paper Review] Learning Channel Inter-dependencies at Multiple Scales on Dense Networks for Face Recognition

Qiangchang Wang, Guodong Guo+1 more|arXiv (Cornell University)|Nov 28, 2017

Face recognition and analysis48 references3 citations

TL;DR

This paper proposes SE-Inception-DenseNet (SEID), a deep learning architecture that integrates multi-scale feature learning, dense layer connections, and Squeeze-and-Excitation (SE) modules to enhance feature representation for unconstrained face recognition. By weighting informative features across multiple scales and dense residual paths, SEID achieves state-of-the-art performance on IJB-A and FaceScrub datasets, with 63.3% verification accuracy at FAR=0.01 and 29.4% at FAR=0.0001, outperforming many existing models despite using moderate training data.

ABSTRACT

We propose a new deep network structure for unconstrained face recognition. The proposed network integrates several key components together in order to characterize complex data distributions, such as in unconstrained face images. Inspired by recent progress in deep networks, we consider some important concepts, including multi-scale feature learning, dense connections of network layers, and weighting different network flows, for building our deep network structure. The developed network is evaluated in unconstrained face matching, showing the capability of learning complex data distributions caused by face images with various qualities.

Motivation & Objective

• To address the challenge of recognizing faces under unconstrained conditions, where image quality, pose, and lighting vary widely.
• To improve feature representation in deep networks by combining multi-scale feature learning with dense connectivity for better gradient flow and parameter efficiency.
• To enhance discriminative power by adaptively weighting feature maps using Squeeze-and-Excitation modules, focusing on informative channels.
• To develop a model that generalizes well on complex data distributions without requiring massive training datasets.

Proposed method

• The proposed SE-Inception-DenseNet (SEID) integrates Inception modules for multi-scale convolutional feature extraction within a DenseNet architecture to enable deep, hierarchical feature learning.
• Dense connections are applied across all preceding layers in each dense block, promoting feature reusability, gradient flow, and regularization.
• Squeeze-and-Excitation (SE) modules are inserted before every transitional and dense block layer to recalibrate channel-wise feature responses based on global context.
• The SE module computes channel-wise attention weights via global average pooling, followed by fully connected layers with ReLU and sigmoid activation to produce learnable scaling factors.
• The model is trained end-to-end on the CASIA-WebFace dataset and evaluated on LFW, IJB-A, and FaceScrub for verification performance.
• The network uses 1×1 convolutions for dimensionality reduction in Inception modules and applies identity shortcuts to stabilize training in deep architectures.

Experimental results

Research questions

• RQ1 Can combining multi-scale feature learning with dense connectivity improve performance in unconstrained face recognition?
• RQ2 How does the integration of Squeeze-and-Excitation modules affect feature representation in deep face recognition networks?
• RQ3 Does the proposed SE-Inception-DenseNet model achieve superior performance on cross-quality face matching compared to state-of-the-art methods?
• RQ4 Can the model maintain high accuracy on standard benchmarks like LFW while generalizing well to diverse image qualities in IJB-A and FaceScrub?

Key findings

• On the IJB-A dataset, SEID achieves 63.3% verification accuracy at FAR=0.01, outperforming FaceNet (25.7%), VGGFace (60.5%), and Light CNN (56.6%) under the same metric.
• At FAR=0.0001 on IJB-A, SEID achieves 26.6% verification accuracy, surpassing SphereFace (24.5%) and Center Loss (16.4%), demonstrating strong robustness to low false accept rates.
• On the FaceScrub dataset, SEID achieves 61.7% verification accuracy at FAR=0.01, exceeding VGGFace (59.5%), Light CNN (50.3%), and Center Loss (49.3%) under the same condition.
• On the LFW dataset, SEID achieves 98.88% verification accuracy when using the SE module before the Inception block, outperforming DeepID2+ (98.95% with 25 networks) and Light CNN (99.33%) under comparable training data.
• The reduction ratio r=4 in the SE module yields the best performance (98.88% on LFW), indicating optimal balance between representational capacity and computational cost.
• SEID shows consistent improvement over baseline Inception and DenseNet models, with a 3.1% gain over Inception (97.6%) and a 3.2% gain over DenseNet (95.7%) on LFW.