Face-sketch learning with human sketch-drawing order enforcement
- PDF / 585,140 Bytes
- 3 Pages / 595 x 842 pts (A4) Page_size
- 60 Downloads / 192 Views
. LETTER .
November 2020, Vol. 63 219103:1–219103:3 https://doi.org/10.1007/s11432-019-2890-8
Face-sketch learning with human sketch-drawing order enforcement Liang CHANG1 , Lihua JIN1 , Lifen WENG2 , Wentao CHAO3 , Xuguang WANG3 , Xiaoming DENG4* & Qiulei DONG5,6,7* 1 School of Artificial Intelligence, Beijing Normal University, Beijing 100875, China; Department of Design Art, Xiamen University of Technology, Xiamen 361024, China; 3 Department of Automation, North China Electric Power University, Baoding 071003, China; 4 Beijing Key Laboratory of Human Computer Interactions, Institute of Software, Chinese Academy of Sciences, Beijing 100190, China; 5 National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China; 6 School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing 100049, China; 7 Center for Excellence in Brain Science and Intelligence Technology, Chinese Academy of Sciences, Beijing 100190, China 2
Received 27 November 2019/Revised 19 January 2020/Accepted 26 February 2020/Published online 12 October 2020 Citation Chang L, Jin L H, Weng L F, et al. Face-sketch learning with human sketch-drawing order enforcement. Sci China Inf Sci, 2020, 63(11): 219103, https://doi.org/10.1007/s11432-019-2890-8
Dear editor, Although face-sketch synthesis generates a sketch from a given face photo automatically [1], it is an open research problem in computer vision [2– 4]. Recently, several deep neural network (DNN) methods for face-sketch synthesis have been proposed with considerable results. However, the knowledge of the human sketch-drawing order is not yet exploited much in the existing DNN methods. Generally, in deep learning, if some intermediate knowledge is explicitly embedded in the DNN layers during learning, the problem of overfitting will be reduced significantly to achieve better performance. In neuroscience, the principle is commonly used in DNN-based sensory cortex modeling [5]. Moreover, by enforcing the DNN training with functional magnetic resonance imaging data, as well as electrophysiological data at different layers, the prediction performance of the trained model is increased significantly. Although our study is in line with these studies, we use the sequential sketching data to enforce the sketch predictions of different DNN layers in the training model. Thus, we propose a new DNN by constraining the face-sketch drawing orders using five key intermediate images in human sketching,
which has the advantage of implicitly embedding the human cognitive knowledge in the DNN-based sketch learning. Because the existing face-sketch datasets do not contain intermediate sketches for a given face photo and cannot support to train our facesketch synthesis network, we present a new orderenforced face-sketch dataset named Ord-Sketch. It is based on the face photos from the CUHK and CUFSF datasets. In our proposed OrdSketch dataset, for each face photo, 25 intermediate sketches with incremental degrees of fineness are sequenti
Data Loading...
