Home

Awesome

SMN: Spectrum-driven Mixed-frequency Network [TMM 2023]

by Peifu Liu, Tingfa Xu, Huan Chen, Shiyun Zhou, Haolin Qin, Jianan Li.

arXiv

TODO

Introduction

Hyperspectral salient object detection (HSOD) aims to detect spectrally salient objects in hyperspectral images (HSIs). However, existing methods inadequately utilize spectral information by either converting HSIs into false-color images or converging neural networks with clustering. We propose a novel approach that fully leverages the spectral characteristics by extracting two distinct frequency components from the spectrum: low-frequency Spectral Saliency and high-frequency Spectral Edge. The Spectral Saliency approximates the region of salient objects, while the Spectral Edge captures edge information of salient objects. These two complementary components, crucial for HSOD, are derived by computing from the inter-layer spectral angular distance of the Gaussian pyramid and the intra-neighborhood spectral angular gradients, respectively. To effectively utilize this dual-frequency information, we introduce a novel lightweight Spectrum-driven Mixed-frequency Network (SMN). SMN incorporates two parameter-free plug-and-play operators, namely Spectral Saliency Generator and Spectral Edge Operator, to extract the Spectral Saliency and Spectral Edge components from the input HSI independently. Subsequently, the Mixed-frequency Attention module, comprised of two frequency-dependent heads, intelligently combines the embedded features of edge and saliency information, resulting in a mixed-frequency feature representation. Furthermore, a saliency-edge-aware decoder progressively scales up the mixed-frequency feature while preserving rich detail and saliency information for accurate salient object prediction. Extensive experiments conducted on the HS-SOD benchmark and our custom dataset HSOD-BIT demonstrate that our SMN outperforms state-of-the-art methods regarding HSOD performance.

<div align="center"> <img src="./pics/RGB-vs-HSI.png" width="450px" height="450px" alt="RGB-vs-HSI"/> </div>

Architecture

<div align="center"> <img src="./pics/SMN.png"/> </div>

Requirements

Our project utilizes Python version 3.8.13, with PyTorch version 1.12.1 and torchvision version 0.13.1. We also use CUDA version 11.3. In addition, our project depends on the following libraries: NATTEN, einops, timm, mmcv, h5py, tqdm.

Please ensure you have these libraries installed to properly run our code.

Checkpoints and Results

Checkpoints, test results, and ground truth edge maps on the HSOD-BIT dataset are available for download from Google Drive or Baidu Drive (XX6K).

Checkpoints, test results, and ground truth edge maps on the HS-SOD dataset are available for download from Google Drive or Baidu Drive (gc81).

Getting Started

Data Preparation

Take HS-SOD as an example. Please download the HS-SOD dataset from https://github.com/gistairc/HS-SOD. After downloading, divide the dataset into training and testing sets and place them in the ./DataStorage/HS-SOD/ directory. The ground truth images of the edges are generated using pidinet. Please place them in the edge_GT/ folder.

If you want to use our checkpoints directly, please download the checkpoints from the link above and place them in the checkpoints/ folder.

The structure of the DataStorage/ should be as follows:

|DataStorage/
    |--- HSOD-BIT/
    |--- HS-SOD/
        |--- test.txt
        |--- train.txt
        |--- checkpoints/
            |--- pvt_v2_b1/
                |--- model-best
            |--- resnet18/
                |--- model-best
            |--- swin_t/
                |--- model-best
        |--- GT/
            |--- train/
                |--- 0001.jpg
                |--- 0003.jpg
                |--- ...
            |--- test/
                |--- 0002.jpg
                |--- 0009.jpg
                |--- ...
        |--- hyperspectral/
            |--- train/
                |--- 0001.mat
                |--- 0003.mat
                |--- ...
            |--- test/
                |--- 0002.mat
                |--- 0009.mat
                |--- ...
        |--- color/
            |--- train/
                |--- 0001.jpg
                |--- 0003.jpg
                |--- ...
            |--- test/
                |--- 0002.jpg
                |--- 0009.jpg
                |--- ...
        |--- edge_GT/
            |--- 0001.jpg
            |--- 0003.jpg
            |--- ...

Testing

To test our model, please unsure that the checkpoints are placed in the checkpoints/ folder. Then, run the following command (take the PVT_v2_b1 backbone as an example):

CUDA_VISIBLE_DEVICES="0" python test.py --backbone="pvt_v2_b1" --model_path="DataStorage/HS-SOD/checkpoints/pvt_v2_b1/model-best" --data_path="DataStorage/HS-SOD" --dataset="HS-SOD"

Training

To train our model, execute the HSOD_run.sh script. This script handles the training process automatically. Model checkpoints will be stored in the checkpoints/ directory. After training, the script will proceed to test the model and save the results in the exp_results/ folder. An automatic evaluation of the model will follow the testing phase.

sh HSOD_run.sh

Please Note: Due to the large size of HSIs, frequent reading can lead to slow training speed. Our approach involves loading HSIs once. The extraction of Spectral Edge and Spectral Saliency is integrated into the DataLoader, and the results are stored in memory. This significantly reduces runtime without substantially increasing memory usage.

BibTeX

@ARTICLE{10313066,
  author={Liu, Peifu and Xu, Tingfa and Chen, Huan and Zhou, Shiyun and Qin, Haolin and Li, Jianan},
  journal={IEEE Transactions on Multimedia}, 
  title={Spectrum-Driven Mixed-Frequency Network for Hyperspectral Salient Object Detection}, 
  year={2024},
  volume={26},
  number={},
  pages={5296-5310},
  keywords={Image edge detection;Feature extraction;Object detection;Hyperspectral imaging;Head;Neural networks;Image color analysis;Hyperspectral salient object detection (HSOD);mixed-frequency attention;spectrum},
  doi={10.1109/TMM.2023.3331196}}

Acknowledgement

Our implementation is mainly based on the following codebases. We gratefully thank the authors for their wonderful works.

NATTEN, ITTI-saliency-detection, Evaluate-SOD

Other Works

If you are interested in our work, please check out our other projects: HSOD-BIT, DMSSN.

License

This project is released under the MIT license.