Amazon Technologies, Inc.

Rajeev Ranjan

Manoj Aggarwal

Gerard Guy Medioni

Dilip Kumar

Prithviraj Banerjee

Korwin Jon Smith

Matthew Christopher Smith

Abhinav Kashyap

Douglas Andrew Hungarter

Baomin Wang

Lavisha Aggarwal

Nichole Stockman

Joshua Adam Cowan

Jared Corso

Joshua Engelsma

Jiuhong Xiao

A biometric identification system uses inputs of a user that are acquired using different modalities. A model having an intersection branch is trained to determine an embedding using features present in two or more modalities (an intersection of modalities). A first input acquired using a first modality is processed by the model to determine a first embedding. A second input acquired using a second modality is also processed by the model to determine a second embedding. Comparison of the first embedding and the second embedding may be used to determine if the first input and the second input are associated with the same user.

System for multi-modality mismatch determination

During enrollment to a biometric identification system an image of at least a portion of a user is acquired and processed to determine a compact representation (CR). The CR may be stored for later comparison to identify the user, to train embedding models, and so forth. The image is not stored. A machine learning system that comprises a variational autoencoder is trained to produce the CR with a loss function that includes a distortion loss, an embedding distance loss, and in some implementations a bitrate loss. The trained encoder is used to determine the CR from an input image, while the decoder is not stored. The CR contains sufficient information to be used as training data for embedding models, instead of acquired images. Training may be more computationally efficient using the CR. Recognition comparisons may be more efficiently performed using the CR, compared to query images.

System to determine compact representation of data

A biometric identification system acquires a multimodal image. The system is trained to determine three embeddings: a first embedding associated with features present in a first modality, an intersection embedding associated with features present in both the first and a second modality, and an XOR embedding associated with features that are not shared between the first modality and the second modality. Once trained, the system may process a multimodal query image to determine query embedding data. Additional information, such as minutiae depicted in the multimodal query image may also be determined. The query embedding data, and in some implementations the additional information, may be compared with enrolled embedding data associated with a previously enrolled user. If the comparison exceeds a threshold value, an identity may be asserted.

System for single image multimodal biometric recognition

This disclosure describes techniques for providing instructions when receiving biometric data associated with a user. For instance, a device may detect a portion of a user, such as a hand. The device may then display a first graphical element indicating a target location for placing the portion of the user above the user-recognition device. Additionally, the device may determine locations of the portion of the user above the device. The device may then display a second graphical element indicating the locations, such as when the locations are not proximate to the target location. Additionally, the device may display instructions for moving the portion of the user to the target location. Based on detecting that the location of the portion of the user is proximate to the target location, the device may send data representing the portion of the user to a remote system.

Electronic device for automated user identification

A neural network is trained to process input data and generate a classification value that characterizes the input with respect to an ordered continuum of classes. For example, the input data may comprise an image and the classification value may be indicative of a quality of the image. The ordered continuum of classes may represent classes of quality of the image ranging from 'worst', “bad”, “normal”, “good”, to “best”. During training, loss values are determined using an ordered classification loss function. The ordered classification loss function maintains monotonicity in the loss values that corresponds to placement in the continuum. For example, the classification value for a “bad” image will be less than the classification value indicative of a “best” image. The classification value may be used for subsequent processing. For example, biometric input data may be required to have a minimum classification value for further processing.

System for training neural network using ordered classes

Biometric input, such as an image of a hand, may be processed to determine embedding vector data that may be used to identify users. Accuracy of the identification is improved by using high resolution inputs to a deep convolutional neural network (DCNN) that is trained to generate the embedding vector data that is representative of features in the input. Training data sets are expensive to develop and thus may be relatively small. During training of the DCNN, confidence loss values corresponding to the entire input as well as particular patches or portions of the input are determined. These patch-wise confidence loss values mitigate potential overfitting during training of the DCNN and improve overall performance of the trained DCNN to determine embedding vector data suitable for identification.

System for training embedding network

A biometric identification system uses inputs acquired using different modalities. A model having an intersection branch and an XOR branch is trained to determine an embedding using features present in all modalities (an intersection of modalities), and features that are distinctive to each modality (an XOR of that modality relative to the other modality(s)). During training, a first loss function is used to determine a first loss value with respect to the branches. Probability distributions are determined for the output from the branches, corresponding to the intersection and XORs of each modality. A second loss function uses these probability distributions to determine a second loss value. A total loss function for training the model may be a sum of the first loss and the second loss. Once trained, the model may process query inputs to determine embedding data for comparison with embedding data of a previously enrolled user.

Growing community of inventors

Seattle, WA, United States of America

Rajeev Ranjan