空间金字塔与局部感受野相结合的相关熵极限学习机

针对空间金字塔词袋模型中空间特征分布信息利用效率低，各类特征融合不充分的问题，该文提出空间金字塔与局部感受野相结合的相关熵极限学习机(SR-CELM)。在特征提取部分，利用多尺度局部感受野对生成的多层级的字典特征分布图进行卷积，并引入局部位置特征和全局轮廓特征。在特征分类部分，提出一种新的网络以融合各部分特征。同时在传统极限学习机训练方法的基础上利用相关熵准则构建判别性约束，推导出权重更新公式以求解网络的输出权重。为验证SR-CELM的有效性，该文分别在数据库Caltech 101, MSRC和15 Scene上进行实验。实验表明SR-CELM能够充分利用特征中可辨识信息，提高分类正确率。

Correntropy Extreme Learning Machine Based on Spatial Pyramid Matching and Local Receptive Field

Considering the problems of inefficient use of spatial information between features and inadequate fusion of different features, a Correntropy Extreme Learning Machine based on Spatial pyramid matching and local Receptive field(SR-CELM) is proposed. In feature extraction part, multi-scale local receptive fields are used to convolve the generated multi-level dictionary feature distribution map, and local position features and global contour features are introduced. In feature classification part, a new network is proposed to fuse the features of each part. Based on the traditional extreme learning machine training method, a discriminative constraint is constructed by using the relevant entropy criterion, and the weight update formula is used to solve the output weight of the new network. In order to verify the effectiveness of the SR-CELM, experiments are performed on the databases Caltech 101, MSRC and 15 Scene. The experiments show that SR-CELM can make full use of the identifiable information in the features and improve the classification accuracy.