A Positive Feedback Spatial-Spectral Correlation Network Based on Spectral Slice for Hyperspectral Image Classification

Shi, Cuiping; Wu, Haiyang; Wang, Liguo

doi:10.1109/tgrs.2023.3242776

Cited by 8 publications

(4 citation statements)

References 60 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Accordingly, it is worthwhile to design a unified network for the same target or sensor. Additionally, more attention is still needed to study the spatial and spectral correlation in HSIs, as motivated by the encouraging performance in various tasks, such as super-resolution mapping [47], super resolution [48] and image classification [49].…”

Section: Discussionmentioning

confidence: 99%

Hyperspectral Target Detection via Global Spatial–Spectral Attention Network and Background Suppression

Wang,

Wang

et al. 2023

IEEE J. Sel. Top. Appl. Earth Observations Remote Sensing

Self Cite

View full text Add to dashboard Cite

The accuracy of hyperspectral target detection (HTD) is often affected by the problems of spectral variation and complex background distribution. Inspired by the powerful representational ability of deep learning, we proposed a 3-D convolution-based global spatial-spectral attention network (GS 2 A-Net) to deal with spectral variation in hyperspectral images (HSIs). GS 2 A-Net uses 3-D convolution kernels of different sizes to capture local spatial and spectral features to achieve multi-scale information interaction. Different from the previous 2-D attention mechanisms, GS 2 A-Net simultaneously considers the information in the spatial and spectral dimensions, and creates a weight map consistent with the size of the original HSI. Furthermore, we proposed a new background suppression strategy based on spectral angle mapping (SAM) to achieve more accurate target detection, which can preserve the targets as much as possible when suppressing the background. The method was validated through experiments on five real-world HSI datasets. Compared with several classical and deep learning-based methods, the proposed method exhibits greater detection accuracy.  Index Terms-Hyperspectral target detection, spectral variation, global spatial-spectral attention network, background suppression. I. INTRODUCTION Hyperspectral sensors record hyperspectral images (HSIs) with hundreds of continuous and narrow bands, reaching a spectral resolution of around 10 nm. Due to its powerful representational ability, HSIs have been widely used in land cover classification [1], [2], target detection [3], [4], anomaly detection [5], [6], image unmixing [7], [8] and change detection [9]. Hyperspectral target detection (HTD) is one of the most challenging issues in these applications, due to the limited amount of known information about target spectra and the background. Indeed, HTD can be regarded as a problem of weakly supervised binary classification.

show abstract

Section: Discussionmentioning

confidence: 99%

Hyperspectral Target Detection via Global Spatial–Spectral Attention Network and Background Suppression

Wang,

Wang

et al. 2023

IEEE J. Sel. Top. Appl. Earth Observations Remote Sensing

Self Cite

View full text Add to dashboard Cite

show abstract

“…However, the mixing operation with randomness is easy to weaken the representation ability (Guo, Mao, and Zhang 2019;Chou et al 2020). This is primarily due to the direct interpolation without considering the complementarity of two features and the neglect of specific feature channels (Hou, Liu, and Wang 2017;Shi, Wu, and Wang 2023;Luo, Xu, and Xu 2022;Zhu et al 2023), which in turn impacts the distribution of prediction results. As illustrated in Figure 1, given a novel sample of "Retriever" and another randomly picked out sample "Linnet", the manifold regularization methods, e.g., Mixup (Zhang et al 2018), CutMix (Yun et al 2019), and PatchMix (Liu et al 2021), interpolate their images and labels to train the classifier for predicting both categories.…”

Section: Introductionmentioning

confidence: 99%

“…This purposeful selection, in contrast to the random selection in manifold regularization, enables the classifier to better concentrate more effectively on the novel content during the training stage. Besides, we also exploit the feature complementarity from similar categories and the discriminability of specific feature channels, which can both provide distinctive patterns for classification (Liu et al 2019;Shi, Wu, and Wang 2023). Building on the above analysis, we propose two attention-based calculations at the instance and channel levels, respectively.…”

Section: Introductionmentioning

confidence: 99%

Boosting Few-Shot Learning via Attentive Feature Regularization

Zhu,

Wang,

et al. 2024

AAAI

View full text Add to dashboard Cite

Few-shot learning (FSL) based on manifold regularization aims to improve the recognition capacity of novel objects with limited training samples by mixing two samples from different categories with a blending factor. However, this mixing operation weakens the feature representation due to the linear interpolation and the overlooking of the importance of specific channels. To solve these issues, this paper proposes attentive feature regularization (AFR) which aims to improve the feature representativeness and discriminability. In our approach, we first calculate the relations between different categories of semantic labels to pick out the related features used for regularization. Then, we design two attention-based calculations at both the instance and channel levels. These calculations enable the regularization procedure to focus on two crucial aspects: the feature complementarity through adaptive interpolation in related categories and the emphasis on specific feature channels. Finally, we combine these regularization strategies to significantly improve the classifier performance. Empirical studies on several popular FSL benchmarks demonstrate the effectiveness of AFR, which improves the recognition accuracy of novel categories without the need to retrain any feature extractor, especially in the 1-shot setting. Furthermore, the proposed AFR can seamlessly integrate into other FSL methods to improve classification performance.

show abstract

“…CNN often includes deep convolutional layers and fixed filter weights, which result in a constrained receptive field and a notable increase in computing costs. Furthermore, the majority of CNN-based classification techniques concentrate on obtaining abstract spatial-spectral feature representations; nevertheless, their capacity to extract deep semantic features is restricted [27][28][29].…”

Section: Introductionmentioning

confidence: 99%

Hyperspectral image classification using a large selective kernel network hybridized to-kenization transformer

Wan,

He,

Gao

et al. 2024

Preprint

View full text Add to dashboard Cite

One of the very popular deep learning-based techniques for classifying hyperspectral image (HSI) is convolutional neural networks (CNNs). However, extensive spatial-spectral infor-mation with several continuous bands is one of HSIs' main characteristics; this invariably leads to issues with large computing costs and computational complexity for network opti-mization. Additionally, current CNN-based classification techniques are only partially able to extract deep semantic characteristics. In response to these problems, this paper presents a novel HSI classification framework, a large selective kernel network, associated with a to-kenization transformer, namely LSKTT, which mainly includes three parts: First, HSI's spectral dimensions are decreased via principal component analysis, and a model for char-acterizing abstract feature representations by cascading a 3-D convolution layer and a 2-D convolution layer is employed. Second, to better model the ranging context of various ob-jects, the large selective kernel network is utilized by modifying its vast spatial receptive field dynamically. Third, a feature tokenization transformer is further utilized to exploit deeper semantic characteristics by converting and learning abstract spatial-spectral features. Experiments with three genuine HSIs show that the proposed LSKTT architecture beats many cutting-edge techniques in both qualitative and quantitative performance.

show abstract

A Positive Feedback Spatial-Spectral Correlation Network Based on Spectral Slice for Hyperspectral Image Classification

Cited by 8 publications

References 60 publications

Hyperspectral Target Detection via Global Spatial–Spectral Attention Network and Background Suppression

Hyperspectral Target Detection via Global Spatial–Spectral Attention Network and Background Suppression

Boosting Few-Shot Learning via Attentive Feature Regularization

Hyperspectral image classification using a large selective kernel network hybridized to-kenization transformer

Contact Info

Product

Resources

About