20160413-11 韩先花: Toward Image Representation in Visual Recognition

报告嘉宾：韩先花 主任研究员（日本日本国立产业技术综合研究所）

报告时间：2016年4月13日（周三）晚21：00（北京时间）

报告题目：Toward Image Representation in Visual Recognition

主持人：姬艳丽（电子科技大学）

报告摘要：

Image representation (Feature extraction) is a hot research direction in visual recognition and understanding. The researches of image representation mainly have gone through three periods: 1) the first generation: low-level image representation extracted as the statistics of the uniformly/binary quantized raw features (color intensity, gradient/orientation and small patch) such as color statistics, HOG and LBP-based texture features, and the fusion and selection methods before 2000; 2) the second generation: middle-level image representation by combining the extracted local descriptors of an image, which mainly includes three steps: descriptor extraction, coding and pooling called as BOF model; 3) the second generation: high-level image representation with the end-to-end deep network, where the near top-level output can be considered the HL feature. Based on the active research on image representation, this lecture introduces our following contributions:

(1)  Instead of the coding/pooling procedures in BOF model, we proposed multilinear manifold learning for fusion the large amount of local descriptors from an image;

(2)  We proposed a simple and robust local descriptor based human perception principle, and extracted high-order statistics based on GMM model.

(3)  The primary experiments of food recognition with the high-level features using deep convolution network have been conducted.

参考文献：

[1] Xian-Hua Han, Yen-Wei Chen, Xiang Ruan,“Multilinear Supervised Neighborhood Embedding of Local Descriptor Tensor for Scene/object Recognition”, IEEE Transaction on Image Processing, Vol. 21, No. 3, pp. 1314-1326, 2012.

[2] Xian-Hua Han, Yen-Wei Chen and Gang Xu, “High-order Statistics of Weber Local Descriptors for Image Representation”, IEEE Transaction on Cybernetics, Vol. 45, Issue 6, pp. 1180-1193, 2015.　

[3] Xian-Hua Han, Jian Wang, Gang Xu and Yen-Wei Chen, “High-order Statistics of Micro-Texton for HEp-2 Staining Pattern Classification”, IEEE Transaction on Biomedical Engineering, Vol. 61, No. 8, pp. 2223-2234, 2014.

[4] Xian-Hua Han, Yen-Wei Chen and Gang Xu, “Integration of Spatial and Orientation Contexts in Local Ternary Patterns for HEp-2 Cell Classification”, Pattern Recognition Letters (In press).

报告人简介：