基于认知科学的研究提出一个新颖的计算模型用于物体识别.特征整合理论为计算模型提供了总体路线.基于最大熵原理构建学习过程,获得必要的先验知识构成认知网络.利用认知网络,将底层的图像特征和高层知识捆绑起来.利用条件随机场的基本概念和原理建模捆绑过程.将计算模型应用于现实世界的物体识别,在标准图像库上进行评估,取得了很好的效果.
We propose a new computational model for object recognition based on the vision cognitive findings. Feature integration theory offers the roadmap for our computing model. We construct the learning procedure to acquire necessary pre-knowledge for the recognition network on the basis of the hypothesis-maximum entropy principle. With the recognition network, we can bind the low-level image features and the high-level knowledge. Fundamental concepts and principles of conditional random fields are employed to model the binding process. We apply our model to real object recognition problem and evaluate it on the benchmark image databases to show its satisfactory performance.
[1] Crick F. Functions of the thalamic reticular complex: the searchlight hypothesis[J]. Proc Natl Acad Sci USA, 1984, 81: 4586-4590.
[2] Sejnowski T J. Open questions about computation in cerebral cortex //McClelland J L, Rumelhart D E. Parallel Distributed Processing. Cambridge: MIT Press, 1986: 372-389.
[3] Treisman A, Gelede G. A feature-integration theory of attention[J]. Cognit Psychol, 1980, 12: 97-136.
[4] Gray C M, König P, Engel A K, et al. Oscillatory responses in cat visual cortex exhibit inter-columnar synchronization which reflects global stimulus properties[J]. Nature, 338: 334-337.
[5] Damasio A R. The brain binds entities and events by multiregional activation from convergence zones[J]. Neur Comput, 1989, 1: 123-132.
[6] Treisman A. Feature binding, attention and object perception[J]. Phil Trans R Soc Lond B, 1998, 353: 1295-1306.
[7] Shutton J, Winn J, Rother C, et al. Textonboost for image understanding: multi-class object recognition and segmentation by jointly modeling texture, layout, and context[J]. IJCV, 2009, 81:2-23.
[8] Li T, Kweon I S. A semantic region descriptor for local feature based image classification //ICASSP. 2008.
[9] Lafferty A, McCallum, Pereira F C N. Conditional random fields: Probabilistic models for segmenting and labelling sequence data //Proceedings of the Eighteenth International Conference on Machine Learning (ICML2001). Williams Town, MA, USA, 2001:282-289.
[10] Wallach H. Efficient training of conditional random fields . Edinburgh, UK: University of Edinburgh, 2002.
[11] Sha F, Pereira F. Shallow parsing with conditional random fields //Proceedings of HLT-NAACL. 2003: 213-220N.
[12] Chen S F, Rosenfeld R. A Gaussian prior for smoothing maximum entropy models . Technical Report CMU-CS-99-108, Carnegie Mellon University, 1999.
[13] Byrd R H, Nocedal J, Schnabel R B. Representations of quasi-newton matrices and their use in limited memory methods[J]. Mathematical Programming, 1994, 63:129-156.
[14] Phan X H, Nguyen M L, Nguyen C T. Flex CRFs:flexible conditional random fields . .http://www.jaist.ac.jp/~hieuxuan/flexcrfs/ flexcrfs.html.
[15] Viterbi A J. Error bounds for convolutional codes and an asymptotical optimum decoding algorithm[J]. IEEE Trans Inform Theory, 1967, IT-13: 260-269.
