ID face photos are widely used for identity verification in many real-world applications. In these cases, the face from ID photo is expected to compare with the face obtained from the daily life image. To improve the ID/life face verification performance, the datasets containing various ID face photos and daily life face images are always needed to train an available CNN model for feature extraction. However, it’s really hard to get uniform distributed data for both the ID photos and daily life images. In most face datasets, each subject only contains one ID photo (for example, obtained from the identity card) while contains varies of life images. This imbalanced distribution causes the difficulty of learning a general deep feature representation for both the ID and life faces, since the daily life images tend to have a higher impact on the feature learning than the ID photos. To address this challenge, we propose a weighted center loss which aims to learn a center of deep features for each class and penalizes the distance between the image features and their corresponding class centers with different weights. By emphasizing the ID photo features in the loss computation, our weighted center loss can train a general CNN which obviously narrows the distance between the ID and life faces. Final experiments demonstrate that with the joint supervision of softmax and weighted center loss, our proposed framework can minimize the intra-class variations caused by the imbalanced ID/life data distribution and significantly improve the ID face based verification accuracy.
Hash coding is a widely used technique in approximate nearest neighbor (ANN) search, especially in document search and multimedia (such as image and video) retrieval. Based on the difference of distance measurement, hash methods are generally classified into two categories: Hamming hashing and Manhattan hashing. Benefitting from better neighborhood structure preservation, Manhattan hashing methods outperform earlier methods in search effectiveness. However, due to using decimal arithmetic operations instead of bit operations, Manhattan hashing becomes a more time-consuming process, which significantly decreases the whole search efficiency. To solve this problem, we present an intuitive hash scheme which uses Flat Binary Code (FBC) to encode the data points. As a result, the decimal arithmetic used in previous Manhattan hashing can be replaced by more efficient XOR operator. The final experiments show that with a reasonable memory space growth, our FBC speeds up more than 80% averagely without any search accuracy loss when comparing to the state-of-art Manhattan hashing methods.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.