Author Affiliations
1Key Laboratory of Photoelectronic Imaging Technology and System of Ministry of Education of China, School of Optics and Photonics, Beijing Institute of Technology, Beijing 100081, China2Beijing Institute of Technology Chongqing Innovation Center, Chongqing 401151, China3Department of Engineering Science, Institute of Biomedical Engineering, University of Oxford, Oxford OX3 7DQ, UK4School of Printing & Packaging Engineering, Beijing Institute of Graphic Communication, Beijing 102600, Chinashow less
Fig. 1. Overall schematic diagram of DL-FSCPHI method.
Fig. 2. The reconstruction strategy proposed in this work. F2 is the measured full-Stokes images, while G2 is the detected polarization-compressed images, containing N2 targets, Nλ2 spectral bands, and Nx × Ny spatial pixels. The epoch, the batch size, and the learning rate are parameters set for model training. The iepoch and the jbatch refer to training the ith epoch and jth batch. F1 is the full-Stokes images predicted from the detected polarization-compressed images G1, containing N1 targets and Nλ1 spectral bands.
Fig. 3. Measured and reconstructed full-Stokes images of three test targets in 6 spectral bands from 560 nm to 660 nm with an interval of 20 nm. The reconstructed images are marked with the PSNR and the SSIM values.
Fig. 4. PSNR and SSIM values of the reconstructed full-Stokes images of the three test targets in 18 spectral bands ranging from 520 nm to 690 nm at intervals of 10 nm.
Fig. 5. Loss curves of the training models under different settings, including two sets of training parameters (epoch = 20, batch size = 7 and epoch = 40, batch size = 5), two sets of polarization angles (θ = 114°, β = 0° and θ = 27°, β = 0°), and two convolution models (DL-M1 and DL-M2).
θ = 27°, β = 0° | DL-M1 | DL-M2 | TwIST |
---|
Evaluation metrics | Epoch = 20 | Epoch = 40 | Epoch = 20 | Epoch = 40 | Accuracy = 0.005 |
---|
Batch size = 7 | Batch size = 5 | Batch size = 7 | Batch size = 5 |
---|
PSNR/dB | S0 | 37.58 | 38.04 | 38.95 | 38.76 | 29.37 | S1 | 22.17 | 22.62 | 22.06 | 22.27 | 10.96 | S2 | 24.87 | 25.17 | 24.38 | 25.22 | 10.37 | S3 | 32.63 | 33.57 | 31.20 | 32.19 | 9.85 | Average | 29.31 | 29.85 | 29.15 | 29.61 | 15.14 | SSIM | S0 | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 | S10 | 0.80 | 0.82 | 0.80 | 0.81 | 0.52 | S2 | 0.89 | 0.90 | 0.87 | 0.88 | 0.52 | S3 | 0.98 | 0.98 | 0.97 | 0.97 | 0.52 | Average | 0.92 | 0.92 | 0.91 | 0.92 | 0.64 |
|
Table 1. Average PSNR and SSIM Values of the Reconstructed Full-Stokes Images of 7 Test Targets in 18 Spectral Bands under Different Settings, Including Two Sets of Polarization Angles (θ = 114°, β = 0° and θ = 27°, β = 0°), Two Convolution Models and One Traditional Algorithm (DL-M1, DL-M2, and TwIST), and Two Sets of Training Parameters (Epoch = 20, Batch Size = 7 and Epoch = 40, Batch Size = 5)