Low-resolution facial expression recognition is broadly required in the field of video surveillance in various real-world environments, which make it necessary to speed up the algorithm while maintaining high recognition accuracy. Since deep forest has the ability of easily distributable on parallel hardware architectures, we consider it as an efficient approach for low-resolution facial expression recognition. Different from the original multi-Grained Cascade forest (gcForest), which is the first model of deep forest, we propose the texture mapping-based gcForest (TM-gcForest) method. TM-gcForest unifies texture-mapping as the first level of representation learning with gcForest which originally consists of only two steps, namely multi-grained scanning and cascade forest. The added layer is especially designed for low-resolution images, which can extract features focusing on the contour of faces and is also helpful for the classification at the step of multi-grained scanning. Experimental results on three facial expression datasets show that TM-gcForest obtains on par or superior results compared with hand-feature methods, original deep forest model and the DNN model, which validates the effectiveness of TM-gcForest in both lab-controlled and wild environments. Especially, TM-gcForest performs much better than hand-feature methods and original gcForest on low-resolution images, indicating that TM-gcForest is well suited for real-world video surveillance.
Low-Resolution Facial Expression Recognition Based on Texture Mapping-Based GcForest
2020-10-14
763595 byte
Conference paper
Electronic Resource
English
Fuzzy Based Facial Expression Recognition
IEEE | 2008
|Facial expression recognition based on anatomy
British Library Online Contents | 2014
|View and Texture-Independent Facial Expression Recognition in Videos using Dynamic Programming
British Library Conference Proceedings | 2005
|Facial Expression Recognition Based on LBP-EHMM
IEEE | 2008
|