Although biometric ear recognition has recently gained a considerable degree of attention, it remains difficult to use currently available ear databases because most of them are constrained. Here, the authors introduce a novel architecture called ScoreNet for unconstrained ear recognition. The ScoreNet architecture combines a modality pool with a fusion learning approach based on deep cascade score-level fusion. Hand-crafted and deep learning methods can be used together under the ScoreNet architecture. The proposed method represents the first automated fusion learning approach and is also compatible with parallel processing. The authors evaluated ScoreNet using the Unconstrained Ear Recognition Challenge Database, which is widely considered to be the most difficult database for evaluating ear recognition developed to date, and found that ScoreNet outperformed all other previously reported methods and achieved state-of-the-art accuracy.

Age estimation from facial images is an important application of biometrics. In contrast to other facial variations like occlusions, illumination, misalignment and facial expressions, ageing variation is affected by human genes, environment, lifestyle and health which make age estimation a challenging task. In this study, the authors propose a new age estimation system which exploits multi-stage features from a generic feature extractor, a trained convolutional neural network (CNN), and precisely combined these features with a selection of age-related handcrafted features. This method utilises a decision-level fusion of estimated ages by two different approaches; the first one uses feature-level fusion of different handcrafted local feature descriptors for wrinkle, skin and facial component, while the second one uses score-level fusion of different feature layers of a CNN for its age estimation. Experiments on the publicly available MORPH-Album-2 and FG-NET databases prove the effectiveness of the novel method. Moreover, an additional experimental study on AgeDB database demonstrates that the proposed method is comparable with the best state-of-the-art system for age estimation using in-the-wild age databases.

Human gait or walking manner is a biometric feature that allows identification of a person when other biometric features such as the face or iris are not visible. In this study, the authors present a new pose-based convolutional neural network model for gait recognition. Unlike many methods that consider the full-height silhouette of a moving person, they consider the motion of points in the areas around human joints. To extract motion information, they estimate the optical flow between consecutive frames. They propose a deep convolutional model that computes pose-based gait descriptors. They compare different network architectures and aggregation methods and experimentally assess various body parts to determine which are the most important for gait recognition. In addition, they investigate the generalisation ability of the developed algorithms by transferring them between datasets. The results of these experiments show that their approach outperforms state-of-the-art methods.

Handwritten biometric recognition (writer identification) is a process of identifying the author of a given handwriting. This process belongs to behavioural biometric systems. This study presents a gradient-based technique to offline writer identification in Persian documents. In the proposed method, some similar segmented characters were used for feature extraction. These characters were selected based on its abundance in the Persian language. Other main advantages of the proposed method included defining Persian stroke concept based on Persian characteristics, computing statistical features from Persian strokes and identifying writer by using only one stroke. The suggested method utilised gradient descriptor to extract three energy-based and eight angle-based features. This feature vector was augmented by averaging and a codebook, which utilised augmented feature vectors, was assigned to each writer for each stroke. For identification, a comparison was made of new stroke codebook with the codebook of all writers in this stroke using Kullback–Leibler distance. To test the suggested method, some characters of a standard database were manually segmented and labelled. In the meantime, a large Persian handwriting database was collected and labelled. The system was evaluated on the segmented and collected database, and displayed absolutely correct results on many of the strokes.

Contactless palmprint is considered more user convenient than other biometrics due to its acquisition simplicity and less-private nature. Many challenges arise which affect the performance of common contact-based methods when applied to a contactless environment. For example, pose and illumination variations affect the layout and visibility of palm lines. This study proposes a SIFT-based method with three main modifications from the traditional SIFT. First, the palm regions with no significant lines/wrinkles are masked out to reduce the false features. A region with multi-lines is then described by multi-descriptors rather than a single one. Second, only query and target keypoints with small rotation difference are compared together, instead of comparing them all. This speed-up the comparison and enhance the accuracy, versus SIFT, by reducing the wrong matches. Third, an align-based refinement is applied to filter out the incorrect matches. The method is tested on three contactless hand databases; IITD, GPDS and Sfax-Miracl. It achieves a verification equal error rate of 0.72, 0.84 and 1.14% and a correct identification rate of 98.9, 99 and 98.9% on each database, respectively. These results are significantly better than the state-of-art methods on the same databases by 1.9% for verification and 3.2% for identification.

Dorsal hand vein recognition is a kind of biometric technique that has emerged in the last two decades. Owing to its safety, accuracy, and effectiveness, more and more researchers are involved in the study. Here, the authors presented a dorsal hand vein recognition system under uncontrolled environments based on biometric graph matching (BGM). Firstly, the authors establish two hand vein databases under natural indoor lighting conditions, i.e. XJTU-A and XJTU-B, with the hand not fixed. Secondly, the authors focus on optimising the image preprocessing steps in terms of region of interest (ROI) extraction, vein segmentation, and vein skeleton extraction. An ‘open’ operation with a large parameter is carried out to make the ROI extraction more abundant based on the maximum inscribed circle. In vein segmentation, the authors use the curvature point algorithm to better extract the vein skeleton. Thirdly, BGM algorithm is adopted to obtain distance measurements. The authors use single distance measure and multiple distance measures to obtain the threshold for recognition, respectively. Finally, the proposed dorsal hand vein recognition system is tested in three databases, and experiment results show that the improvement of the entire algorithms leads to high accuracy and strong robustness of the recognition system, whether under uncontrolled or controlled conditions.

Dimensional emotion estimation (e.g. arousal and valence) from spontaneous and realistic expressions has drawn increasing commercial attention. However, the application of dimensional emotion estimation technology remains a challenge due to issues such as manual annotation and evaluation. In this work, the authors introduce an automatic annotation and emotion prediction model. The automatic annotation is performed through three main steps: (i) label initialisation, (ii) automatic label annotation, and (iii) label optimisation. The approach has been validated on different language databases with different types of emotion expressions, including spontaneous, acted and induced emotional expressions. Compared with non-optimisation of the predicted labels, the process of optimisation improves the concordance correlation coefficient (CCC) values by an average of 0.104 for arousal and 0.051 for valence. Furthermore, the standard variation between annotated values and the ground truth is reduced to an average of 0.44 for arousal and 0.34 for valence. Finally, the CCC values using the proposed model reach 0.58 for arousal and 0.28 for valence, which further verifies the feasibility and reliability of the proposed model. The proposed method can be used to reduce labour intensive and time-consuming manual annotation work.