Palmprint Recognition Using Realistic Animation Aided Data Augmentation

Abstract

In this paper, a palmprint augmentation algorithm based on 3D animation is proposed for enhancing contactless palmprint recognition performance. Contactless palmprint varies in position, orientation and musculoskeletal deformations. As the existing contactless databases are small, they contain only a few such variations of a palm. Popular data augmentation approaches, including translation, rotation and scaling, have been used to increase the dataset size and its diversity, but these methods do not simulate non-linear deformation of the hand. Some researchers have used 3D and computer graphic techniques to generate more data for training deep networks. These techniques are application-specific. The proposed algorithm makes use of a 3D hand model to simulate muscular and skeletal deformations of the hand. The deformations from the 3D model are applied to 2D palmprint images to generate new palmprint images with the same identities. Four deep networks, Alexnet, VGG-16, Resnet-50 and Inception-V3 and two contactless palmprint databases, IITD and CA-SIA, are employed to evaluate the proposed algorithm. The proposed algorithm is compared with the standard augmentation methods. The experimental results show that the proposed augmentation algorithm reduces EER and Rank-1 error rate.

Evaluating the Efficacy of Alexnet for Detection of Thyroid Cancer

“…In this study, the Digital Database of Thyroid Ultrasound Images (DDTI), a publicly accessible dataset consisting of 347 B-mode thyroid ultrasound images has been utilized. Radiologists have employed the Thyroid Imaging Reporting and Data System (TIRADS) for patient classification [33] [34]. The images underwent cropping to specific dimensions, ensuring a maintained square shape in the cropped region.…”

“…It consists of five convolutional layers and three fully connected layers. The first five layers of AlexNet are convolutional layers, which apply filters to the input image to extract features [33] [34]. These convolutional layers are followed by maxpooling layers, which downsample the feature maps to reduce their spatial dimensions.…”

Synthetic data generator: understanding human face & body via image synthesis

“…We can see that some regions of the original covered depth data (red) are blurred, which will mislead the prediction. Compared to covered depth data, the reconstructed depth images (green) can provide more pose information than the original covered Animating a human with a novel view and expression sequence from a single image opens the door to a wide range of creative applications, such as talking head synthesis [228,229], augmented and virtual reality (AR/VR) [230], image manipulation [4,231,35], as well as data augmentation for training of deep models [232,34,33]. Early works of image animation mostly employed either 2Dbased image generation models [233,234,235,236], or 3D parametric models [237,238,239,240] (e.g., 3DMM [241]), but they mostly suffer from artifacts, 3D inconsistencies or unrealistic visuals.…”