CNN-EFF: CNN Based Edge Feature Fusion in Semantic Image Labelling and Parsing

Srivastava, Vishal; Biswas, Bhaskar

doi:10.1007/s11063-021-10704-6

Cited by 7 publications

(1 citation statement)

References 52 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In the past decade, deep learning-based classification techniques have been applied to HS image classification, effectively extracting spatial-spectral features and achieving high accuracy [4,9]. Advanced deep neural networks, including convolutional neural networks (CNNs) [9,10], recurrent neural networks (RNNs) [11], generative adversarial networks (GANs) [12], and stacked autoencoders [13], have been employed. However, due to their limited ability to model relationships between samples, they fail to consider the long-range spatial relations contained in the HS image pixels during classification.…”

Section: Introductionmentioning

confidence: 99%

HyperGCN - A Multi-layer Graph NeuralNetworks to Enhance Hyperspectral Image Classification

Srivast,

Chaurasia

2023

Preprint

View full text Add to dashboard Cite

Graph neural networks (GNNs) have recently attracted much attention due totheir exceptional performance in a wide range of applications, including hyperspectralimage classification. However, most GNN-based models developed forhyperspectral image classification are shallow-layer models that combine CNNand GNN to improve performance. The use of fewer layers in their GNN models isattributed to the observed degradation in performance as the depth of GNN modelsincreases. This study proposes HyperGCN, an exclusive GNN-based modelwith multiple graph convolutional layers that exploit the rich spectral informationcontained in hyperspectral images and improve classification performance.HyperGCN prevents performance degradation by incorporating over-smoothingresistanttechniques into its architecture. Extensive experiments on widely usedhyperspectral benchmark datasets such as the Indian Pines dataset and thePavia University dataset show that HyperGCN outperforms traditional graphconvolutional neural network models in all performance metrics, including overallaccuracy, average accuracy, class-specific accuracy, and Cohen’s kappa coefficient.HyperGCN also outperformed fusion models that combine the benefits of CNNand GNN when classifying hyperspectral images.

show abstract