Guest Editors' Introduction: Special issue on deep learning with applications to visual representation and analysis

The last decade has witnessed an exponential growth of visual information and an unprecedentedly broad range of applications of image and video. To efficiently access, utilize, and transmit visual information, visual representation and analysis have attracted intensive attention and become one of the most active research topics in the fields of signal processing, computer vision, and artificial intelligence (AI). Recently, deep learning has advanced as an AI approach that can automatically discover good representations and model high-level abstractions from data. It has achieved record-breaking performance on a spectrum of visual analysis tasks. At the same time, the full potential of deep learning for visual representation and analysis has yet to be explored and many theoretical and practical issues remain unsolved. This special issue consists of nine papers selected from the submission, which report new research explorations in employing, improving, and designing deep learning algorithms for visual representation and analysis. As guest editors, we received strong support from a team of committed reviewers and the journal management team, especially EIC Frédéric Dufaux. Two papers of this special issue are about facial image analysis. The paper “Landmark Perturbation-Based Data Augmentation for Unconstrained Face Recognition” written by Jiang jing Lv, Cheng Cheng, Guodong Tian, Xiangdong Zhou, and Xi Zhou aims to improve the robustness of deep convolutional neural networks with respect to facial landmark misalignment. A data augmentation method is proposed to achieve this goal. Given a facial image, this method automatically perturbs landmark positions to generate a large number of misaligned copies to train the networks. Experimental study on multiple benchmark data sets shows that this augmentation method can attain better recognition performance in both face verification and identification cases. The paper “FaceHunter: A Multi-task Convolutional Neural Network Based Face Detector” by Dong Wang, Jing Yang, Jiankang Deng, and Qingshan Liu proposes a new multi-task-based convolutional neural network for face detection. Its objective function considers both the discrimination between face and non-face images and the accuracy of bounding box regression, which allows region proposal classification and bounding box refinement to be jointly performed. To improve computational efficiency, regional proposal network is employed to directly generate proposals from convolutional feature map. Experimental result on five face data sets demonstrates the state-of-the-art performance of this detection framework. Pedestrian detection plays a key role in a wide range of practical applications on automotive, surveillance and robotics. A deep