Strengths and Weaknesses of Deep Learning Models for Face Recognition Against Image Degradations (1710.01494v1)

Abstract: Deep convolutional neural networks (CNNs) based approaches are the state-of-the-art in various computer vision tasks, including face recognition. Considerable research effort is currently being directed towards further improving deep CNNs by focusing on more powerful model architectures and better learning techniques. However, studies systematically exploring the strengths and weaknesses of existing deep models for face recognition are still relatively scarce in the literature. In this paper, we try to fill this gap and study the effects of different covariates on the verification performance of four recent deep CNN models using the Labeled Faces in the Wild (LFW) dataset. Specifically, we investigate the influence of covariates related to: image quality -- blur, JPEG compression, occlusion, noise, image brightness, contrast, missing pixels; and model characteristics -- CNN architecture, color information, descriptor computation; and analyze their impact on the face verification performance of AlexNet, VGG-Face, GoogLeNet, and SqueezeNet. Based on comprehensive and rigorous experimentation, we identify the strengths and weaknesses of the deep learning models, and present key areas for potential future research. Our results indicate that high levels of noise, blur, missing pixels, and brightness have a detrimental effect on the verification performance of all models, whereas the impact of contrast changes and compression artifacts is limited. It has been found that the descriptor computation strategy and color information does not have a significant influence on performance.

• The paper demonstrates that image degradations, especially blur and noise, significantly lower verification accuracy in deep learning models.
• It systematically compares four CNN architectures on the LFW dataset, highlighting VGG-Face's superior resilience to noise.
• The study underscores the need for optimized descriptor computation and better handling of incomplete facial data in biometric systems.

Deep Learning Models for Face Recognition: Analyzing Resilience Against Variations

The paper, "Strengths and Weaknesses of Deep Learning Models for Face Recognition Against Image Degradations," explores the robustness of four deep convolutional neural networks (CNNs) architectures—AlexNet, VGG-Face, GoogLeNet, and SqueezeNet—when subjected to various image-quality degradations in the context of face verification tasks. Using the Labeled Faces in the Wild (LFW) dataset, the paper systematically investigates the effects of several covariates such as blur, JPEG compression, noise, brightness, contrast, missing pixels, model architectures, color information, and descriptor computation strategies. With deep CNN models, understanding the variability in performance due to these factors is crucial, given their prominent application in biometric systems.

Image Quality Covariates' Impact

Most notably, the experimentation delineates the detrimental effects of blur, noise, brightness, and missing pixels on model performance, while showing limited impact from contrast changes and JPEG compression artifacts. Specifically, blurring proved to be significantly influential, causing substantial drops in the models’ verification accuracy. Interestingly, AlexNet, SqueezeNet, and GoogLeNet exhibited similar performance degradation patterns under blur conditions, with GoogLeNet being more sensitive than others.

The robustness against JPEG compression errors until extremely low quality settings is indicative of model reliance on critical DCT coefficients. The evaluation with Gaussian and salt-and-pepper noise further emphasized the superiority of VGG-Face in maintaining performance under noisy conditions compared to other model architectures, pointing to its higher resilience to noise perturbations.

Model-Related Covariates' Consequences

While the VGG-Face model consistently demonstrated robustness across various covariates, our analysis recorded exceptions, particularly concerning image brightness variations. Despite deep learning models’ promising performance with color images, grayscale transformations revealed minimal accuracy loss across the studied architectures, indicating fact that existing networks do not fully leverage available color information.

Moreover, evaluation of descriptor computation strategies showed that sampling face image patches for generating descriptors did not universally enhance performance, with only slight improvements noted for VGG-Face and SqueezeNet. This opens discussions on optimizing descriptor computation to enhance verification stability.

Implications and Future Directions

This insightful assessment lays the groundwork for potential advancements in deep learning-based face recognition systems. Key areas highlighted include:

• Image Enhancement Technologies: Developing algorithms to amend the recognition efficacy for images enduring blur or low-resolution challenges.
• Efficient Utilization of Color Information: Given the negligible effect of color versus grayscale assessments, future models could maximize their use of chromatic data or efficiently operate on grayscale setups for reduced complexity.
• Handling Partial Data: Addressing recognition performance declines due to occluded facial regions, particularly around the periocular area, requires attention towards creating models accommodating incomplete data.

In summary, the paper offers a rigorous comparison of popular CNN architectures, spotlighting significant aspects that influence image recognition performance and suggesting areas ripe for further investigative efforts. While existing models show commendable proficiency in face verification tasks, ongoing refinement in dealing with image-quality degradation remains a pressing need for advancing biometric systems. The release of trained models for community use further reinforces the collaborative drive to optimize deep learning technology in face recognition applications.