Hyperspectral Salient Object Detection​



In this work, we proposed a salient object detection model on hyperspectral images by applying manifold ranking (MR) on self-supervised Convolutional Neural Network (CNN) features (high-level features) from unsupervised image segmentation task. Self-supervision of CNN continues until saliency maps converges to a defined error between each iteration. We demonstrated that proposed model can better accuracy compared with the state-of-the-arts hyperspectral saliency models including the original MR based approach. Moreover, we prepared and released a dataset for benchmarking, and proposed model is also tested and compared with existing hyperspectral saliency approaches on it.

  • Hyperspectal Image Dataset
  • CNN
  • Unsupervised Image Representation and Clustering
  • Hyperspectral Salient Object Detection


HS-SOD Dataset and Saliency Detection Model
Hyperspectral Salient Object Detection Dataset ( https://github.com/gistairc/HS-SOD )

Dataset contains:

  • Hyperspectral data: 60 hyperspectral images 768x1024 pixel resolution and 81 spectral channels (range within visible spectrum: 380 nm -720 nm)
  • color: 60 color images of hyperspectral dataset rendered in sRGB for visualization
  • ground-truth salient objects: 60 ground-truth binary images of salient objects manually selected
  • Publications
    • Hyperspectral Image Dataset for Benchmarking on Salient Object Detection (IEEE QoMEX 2018)
    • Salient Object Detection on Hyperspectral Images Using Features Learned from Unsupervised Segmentation Task (IEEE ICASSP 2019)

Proposed Saliency Model with Unsupervised Deep


Visual attention mechanism is an important part of the human visual system enabling us to process large amount of information by focusing on distinctive visual regions to observe and analyze the scene. Moreover, attention process can be task-independent or task-dependent in its nature meaning that it can also help to find out attentional semantic information or low-level distinctive features.

Computational models to try similar behavior tries to generate attention maps so called saliency map in the field of computer vision, which can help to improve many vision task for various sensory information including the data from hyperspectral camera.

Other than the salient object detection, possible applications in hyperspectral imaging can be :

  • Hyperspectral data compression: salient regions can be kept with higher details with less compression and relatively less important areas can be compressed more
  • Hyperspectral image recognition: In supervised learning, some models may use various regional candidate for image representation and feature extraction with random sampling. Therefore, in these case, to avoid potential background information, attention can be used for potential foreground focus.
  • Hyperspectral 3D reconstruction: Redundant spectral information can be helpful for 3D reconstraction of the scene, and hyperspectral salient key points and features can also improve the quality.
研究開発プロジェクト NEDO 人工知能技術適用によるスマート社会の実現の成果
研究機関 国立研究開発法人 産業技術総合研究所
主要研究者 Nevrez Imamoglu