Fig. 1. (a) Schematic layout: overview of the snapshot multispectral imaging system, illustrating the arrangement of the circular LED array, the digital micromirror device (DMD), optical components (collimator, diffuser, and lenses), and the monochrome imaging camera. (b) Experimental setup.

Fig. 2. (a) and (b) Visualizations of image reconstruction results of two test image patterns illuminated under different wavelengths. Each column refers to an image with the monochrome defocused sensor input, network output, and multispectral target (ground truth). Pseudo colors represent the corresponding illumination LED color (center wavelength).

Fig. 3. Visualization of image reconstruction results under various illumination configurations. Each column displays the predicted image alongside the corresponding target (ground truth) image for different combinations of activated LEDs that were on.

Fig. 4. Image reconstruction performance analysis. (a) Confusion matrix of spectral channel accuracy. (b) Classification metrics against the number of concurrent LED illuminations: Plots of sensitivity, specificity, and F1 score against the number of concurrent LED illuminations. (c) Energy difference values as a function of the number of concurrent LED illuminations; boxplots depict the distribution of energy differences for various numbers of concurrent illuminations, indicating the shift in the prediction accuracy and variance. (d) Energy difference values as a function of the illumination wavelengths. We used light green to signify “losses” and light blue for “leakages.”

Fig. 5. Distributions of (a) PSNR, (b) RMSE, (c) SSIM, and (d) NMSE metrics for the reconstructed images at different illumination wavelengths.

Fig. 6. (a) Distributions of structural quality metrics for the reconstructed multispectral images with respect to the number of concurrent LED illuminations. (b) Confusion matrices with different numbers of concurrent LED illuminations.

Fig. 7. Reconstruction results on synthetic test data with dual-LED spectral illumination on different patterns. Each test input simulates a monochrome snapshot formed by the superposition of two independent single-channel illuminations on two distinct image patterns. The mFIN-generated spectral reconstructions (left) are compared against the combined ground truth targets (right, in each column).

Fig. 8. Network architecture. (a) mFIN. (b) Dense links: each output tensor of the dSPAF group is appended and fed to the subsequent one. (c) Detailed schematic of dSPAF modules. See Sec. 4 for details.