Reg-Superpixel Guided Convolutional Neural Network of PolSAR Image Classification Based on Feature Selection and Receptive Field Reconstruction

被引:4
作者
Shang, Ronghua [1 ]
Zhu, Keyao [1 ]
Feng, Jie [1 ]
Wang, Chao [2 ]
Jiao, Licheng [1 ]
Xu, Songhua [3 ]
机构
[1] Xidian Univ, Sch Artificial Intelligence, Key Lab Intelligent Percept & Image Understanding, Minist Educ, Xian 710071, Peoples R China
[2] Zhejiang Lab, Res Ctr Big Data Intelligence, Hangzhou 311100, Peoples R China
[3] Xi An Jiao Tong Univ, Afffliated Hosp 2, Inst Med Artiffcial Intelligence, Xian 710004, Peoples R China
基金
中国国家自然科学基金;
关键词
Feature extraction; Convolutional neural networks; Training; Synthetic aperture radar; Image reconstruction; Coherence; Neural networks; Feature selection; image classification; polarimetric decomposition; receptive field remodeling; DECOMPOSITION; MODEL;
D O I
10.1109/JSTARS.2023.3268177
中图分类号
TM [电工技术]; TN [电子技术、通信技术];
学科分类号
0808 ; 0809 ;
摘要
The convolutional neural network (CNN) has a poor performance in nonuniform and edge regions due to the limitation of fixed receptive field. At the same time, feature stacking of input data can bring burden and overfitting to the network. To solve these problems, this article proposes a reg-superpixel guided CNN based on feature selection and receptive field reconstruction. First, a feature selection method is designed, which uses polarimetric SAR statistical distribution features to calculate distance and similarity, and selects features that are easier to identify to avoid the negative impact of low distinguishing features on classification. Second, the reg-superpixel, which means regular superpixel, is used to reconstruct the receptive field and represent the features of the central pixel. The classification result of the central pixel is extended to the whole superpixel during the test. This method can extend the pixel-level CNN network to superpixel-level. Finally, by using edge information of the small-scale superpixel and spatial information of the large-scale superpixel to adjust receptive field of the central pixel, the classification results with uniform smooth region and high edge fitting can be generated. Experimental results with four state-of-the-art methods on four datasets show that feature selection and multiscale reg-superpixel network is effective for polarimetric SAR classification problems.
引用
收藏
页码:4312 / 4327
页数:16
相关论文
共 35 条