Low-sample size remote sensing image recognition based on a multihead attention integration network
- PDF / 1,681,252 Bytes
- 16 Pages / 439.37 x 666.142 pts Page_size
- 58 Downloads / 182 Views
Low-sample size remote sensing image recognition based on a multihead attention integration network Zesong Wang 1,2
2
& Cui Zou & Xianping Cui
2
Received: 21 February 2020 / Revised: 27 July 2020 / Accepted: 13 August 2020 # Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract
For a long time, small sample recognition for hyperspectral images has been a popular research topic. It is very difficult for an algorithm to simultaneously satisfy the requirements of feature mining, feature selection and feature integration. The traditional single model has difficulty completing multiple tasks at the same time, ultimately leading to poor small sample recognition results for remote sensing images. This paper proposes a multimodel joint algorithm for deep feature mining based on multiscale convolution (MC) under multihead attention (MA) and deep feature integration based on bidirectional independent recurrent neural networks (BiIndRNNs), MACBINet. First, this paper proposes a multihead attention mechanism that assigns multiple weight coefficients to each feature to better select remote sensing image features; then, it implements the deep mining of features and the retention of multiple deep features through multiscale convolution. Subsequently, it implements contextual semantic information integration for longsequence features through bidirectional independent recurrent neural networks to avoid the problem of gradient disappearance during training on a small sample of data. Finally, the softmax function is used to perform recognition on three public remote sensing data sets. The experimental results prove that our proposed MACBINet achieves the best results to date for small sample classification. Keywords Multihead attention mechanism . Multiscale convolution . Bidirectional independent recurrent neural network
* Zesong Wang [email protected] Cui Zou [email protected] Xianping Cui [email protected]
1
Qingdao Huanghai University, Qingdao 266427, China
2
Big Data Institute, Qingdao Huanghai University, Qingdao 266427, China
Multimedia Tools and Applications
1 Introduction Because hyperspectral images are composed of hundreds of bands, each pixel in a hyperspectral image contains rich information. Hyperspectral images are widely used for various purposes, including accurate urban distinction [31, 36], land type segmentation [2, 3], and marine object recognition [15, 38]. However, the large number of high-dimensional bands causes HSIs to contain many redundant features, introducing difficulties and long processing times for image recognition. Recently, researchers have developed new methods for feature selection. Cui J et al. [10] have proposed a novel WA-BS feature subset selection algorithm, which solves the problem of inaccurate temporal and spatial information for remote sensing images. Compared with ReliefF, mRMR and LeastC, WA-BS achieves a higher classification accuracy. Chen T et al. [9] developed a sparse regularized feature learning method to select and discriminate features of remote
Data Loading...
