Low Resolution Face Recognition Using a Two-Branch Deep Convolutional Neural Network Architecture (1706.06247v1)

Abstract: We propose a novel couple mappings method for low resolution face recognition using deep convolutional neural networks (DCNNs). The proposed architecture consists of two branches of DCNNs to map the high and low resolution face images into a common space with nonlinear transformations. The branch corresponding to transformation of high resolution images consists of 14 layers and the other branch which maps the low resolution face images to the common space includes a 5-layer super-resolution network connected to a 14-layer network. The distance between the features of corresponding high and low resolution images are backpropagated to train the networks. Our proposed method is evaluated on FERET data set and compared with state-of-the-art competing methods. Our extensive experimental results show that the proposed method significantly improves the recognition performance especially for very low resolution probe face images (11.4% improvement in recognition accuracy). Furthermore, it can reconstruct a high resolution image from its corresponding low resolution probe image which is comparable with state-of-the-art super-resolution methods in terms of visual quality.

• The paper proposes a dual-branch deep CNN that minimizes the Euclidean distance between LR and HR facial features through coupled, non-linear mappings.
• Its LR branch uses a 5-layer super-resolution network combined with a 14-layer HR branch, achieving an 11.4% accuracy improvement on extremely low-resolution images.
• The method enhances face recognition under varied conditions and enables high-quality HR reconstruction from LR inputs for practical surveillance applications.

Low Resolution Face Recognition Using a Two-Branch Deep Convolutional Neural Network Architecture

This paper introduces a dual-branch deep convolutional neural network (DCNN) approach to enhance low-resolution face recognition. The authors propose a coupled mappings technique where two branches of DCNNs map low resolution (LR) and high resolution (HR) face images to a shared common space via non-linear transformations. The HR branch contains 14 layers, encompassing convolutional and fully connected layers akin to VGGNet configurations, while the LR branch includes a 5-layer super-resolution network followed by the same 14-layer network as the HR branch. The objective is to minimize the Euclidean distance between corresponding HR and LR features in the common space, which is achieved through backpropagation.

Methodology and Results

The research evaluates the proposed method on the FERET dataset, notable for its variety in expression, illumination, and aging conditions. Recognition accuracy sees an 11.4% improvement in extremely low-resolution conditions (such as $6 \times 6$ pixels) over competitive state-of-the-art techniques, indicating the robustness of the proposed architecture in resolving LR face images. Additionally, this method also facilitates the reconstruction of HR images from their LR counterparts with visual quality comparable to existing super-resolution methods. The comprehensive evaluation demonstrates a significant improvement in face recognition across various probe resolutions and conditions, such as illumination and aging.

Implications and Future Directions

The implications of this two-branch DCNN architecture are multifaceted. Firstly, the significant improvement in recognition accuracy for very low-resolution images can revolutionize applications in surveillance systems where high-quality facial images are often unavailable. The ability to reconstruct HR images from LR inputs extends practical usability, serving both recognition and visualization purposes.

The theoretical contributions lie in leveraging non-linear mappings to transmute lower-quality features into a domain where they become more linearly distinguishable relative to HR images. This method stands apart by training a coupled non-linear DCNN framework to optimize both feature extraction and image resolution simultaneously, something prior methodologies did not fully explore.

Future directions could explore extending this architecture to multi-view face recognition scenarios, where resolution and angle variances are even more pronounced. Furthermore, examining the adaptability of the architecture across diverse populations and real-world conditions, including occlusions, could further the application of this method. Research into reducing computational load without sacrificing accuracy could also enhance the deployment feasibility on devices with limited processing capabilities.

Overall, this paper lays a crucial foundation for face recognition technology, particularly in constrained environments, by bridging the gap between observational constraints (i.e., low-resolution imaging) and system capabilities within the field of neural network architectures.