基于神经认知计算模型的高分辨率遥感图像场景分类

通过大脑对外界环境感知的神经结构与认知功能的相关研究,构建仿脑的媒体神经认知计算(multimedia neural cognitive computing, MNCC)模型。该模型模拟了感官的信息感知、新皮层功能柱的认知功能、丘脑的注意控制结构、海马体的记忆存储和边缘系统的情绪控制环路等大脑基本的神经结构和认知功能。在此基础上,构建基于MNCC的高分辨率遥感图像场景分类算法。首先,图像经仿射变换后切分为若干图块,通过深度神经网络提取图块的稀疏激活特征,采用概率主题模型获取图块初始场景类别,并利用图块分类错误信息反馈控制场景显著区特征的提取;其次,根据图块的上下文获取场景语义的时空特征,并在此基础上进行图块分类和场景预分类;最后,用场景预分类误差构造奖惩函数,控制和选择深度神经网络中场景区分度较大的稀疏激活特征,并通过增量式强化集成学习,获得最后的场景分类。在两个标准的高分辨率遥感图像数据集上的实验结果表明,MNCC算法具备较好场景分类结果。

Scene classification of high-resolution remote sensing image based on multimedia neural cognitive computing

According to the related research of the brain of neural structures and cognitive function which apperceive the external environment, the brain-like model of multimedia neural cognitive computing (MNCC) is built. The model simulates the basic neural structures and the cognition of brain, such as the sensory information perception, the cognition of the neocortex column, the attention control structure of the thalamus, the hippocampus memory and emotional control circuits of the limbic system, and a scene classification MNCC-based algorithm for high-resolution remote sensing images is established. Firstly, the algorithm extracts sparse activation features of the deep neural network from the sub-blocks image after affine transformation, and gets initial category of the sub-blocks image with the probability topic model, then controls features extraction of the saliency area by sub-block classification error.Secondly, the temporal-spatial features of scene semantic are acquired by sub-blocks context, then sub-blocks categorization and scene pre-classification are processed to obtain initial scene labels. Finally, the scene pre-classification error is used for construction rewards function to control and select the most discrimination sparse activation features of deep neural network, and the final scene label is obtained by the incremental reinforced ensemble learning algorithm. Experiment results show that the MNCC algorithm presented in this paper has better performance of scene classification on the two standard high resolution remote sensing scene datasets.