Kim,Seong TaeChoi,YeoreumRo,Yong ManBrömme,ArslanBusch,ChristophDantcheva,AntitzaRathgeb,ChristianUhl,Andreas2017-09-262017-09-262017978-3-88579-664-0https://dl.gi.de/handle/20.500.12116/4642In the past few decades, automatic face recognition has been an important vision task. In this paper, we exploit the spatial relationships of facial local regions by using a novel deep network. In the proposed method, face is spatially scanned with spatial long short-term memory (LSTM) to encode the spatial correlation of facial regions. Moreover, with facial regions of various scales, the complementary information of the multi-scale facial features is encoded. Experimental results on public database showed that the proposed method outperformed the conventional methods by improving the face recognition accuracy under illumination variation.enFace recognitionfacial feature representationspatial LSTMdeep learning1617-5468