Fig. 1. Principle of the LightguideCam computational imaging system. (a) Optical path for the LightguideCam. Light from the object is split by the lightguide. Some light passes through directly to the user’s eye, while the rest is guided to an image sensor, forming a blurred image due to the system’s spatially varying point spread function (PSF). The deep neural network (DNN) reconstructs the original object from this sensor measurement. (b) General schematic of a computational imaging system, where a DNN is trained to solve the inverse problem of recovering a clean object image from a measurement with artifacts.

Fig. 2. Experimental setup and data acquisition. (a) Photograph of the experimental setup. An OLED display, placed 70 cm from the LightguideCam, projects ground truth images. The LightguideCam consists of a lightguide and an image sensor. A cooling fan is used to stabilize the sensor temperature and reduce thermal noise. (b) Schematic of the dataset collection process. Ground truth images are loaded onto the display, and the corresponding distorted images are captured by the image sensor, forming the paired dataset for network training.

Fig. 3. Reconstruction network architecture and training analysis. (a) The architecture of the physics-informed Multi-Wiener Net. The input measurement b is transformed into the frequency domain and processed by a bank of parallel Wiener filters, which use learnable PSFs (Hi) and regularization parameters (λi). The resulting intermediate reconstructions are transformed back to the image domain and fed into a U-Net, which outputs the final reconstruction v. The network is trained end-to-end by minimizing the mean squared error (MSE) between the reconstruction and the ground truth. (b) Visualization of the network’s deconvolution stage. From left to right: the initial calibrated PSFs, the PSFs after training, and the corresponding intermediate reconstructions. (c) Training and validation loss curves for the R, G, and B color channels, showing convergence over 60 epochs.

Fig. 4. Quantitative and qualitative comparison of reconstruction results. (a)–(c) Violin plots comparing the distribution of peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and correlation coefficient (CC) across the test dataset for the raw measurement, FISTA reconstruction, and our M. W. Net reconstruction. The red bars indicate the mean values. (d) A representative visual comparison, showing from left to right: the raw blurred measurement, the FISTA reconstruction, the M. W. Net reconstruction, and the ground truth image. (e) Corresponding residue images, calculated as the absolute difference between each reconstruction and the ground truth. A darker image signifies a smaller error. (f) Pixel-wise SSIM maps. Warmer colors (yellow/red) indicate higher structural similarity to the ground truth, while cooler colors (blue) indicate lower structural similarity.

Fig. 5. Reconstruction of 3D scenes with varying depths. The top row shows the raw captured measurement, and the bottom row shows the corresponding reconstruction from our network. The network was trained only on images from a single depth plane (70 cm). Consequently, it sharply reconstructs objects at that depth while objects at other discrete depths or parts of a continuous object outside the focal plane remain blurred.