This page delves into further hardware advancements to achieve optimal PSF sparsity, and enhanced image reconstruction algorithmic techniques.
1. PSF Refinement
The diffuser mask plays a pivotal role in the hardware design, as different masks yield varying PSF patterns directly impacting reconstruction performance. Therefore, achieving a sharp, depth-dependent PSF is crucial. Despite efforts in this capstone project, a similar PSF to the DiffuserCam PSF was not attained using double-sided Scotch tape. Future endeavors aim to obtain a high-contrast PSF by experimentally subtracting background and implementing high-pass filtering. Additionally, exploring alternative diffusing materials, such as plastic bags or microlens arrays, holds promise for PSF refinement.
2. Algorithm Speed Enhancement
The current algorithm employed is basic gradient descent without momentum parameters. Enhancing algorithmic efficiency is imperative for system improvement. While the current interface boasts user-friendliness, the lack of packaged algorithms hampers students’ learning experiences in optimization. To expedite algorithmic processes, incorporating momentum parameters is recommended. Furthermore, implementing the algorithm in PyTorch and executing gradient descent on GPUs could significantly enhance speed, particularly with large datasets.
3. 3D Printing Integration
Our hardware setup is on a benchtop optical table, limiting portability. For educational applicability, transitioning to a portable camera configuration is envisioned. This entails utilizing a smaller sensor with reduced pixel counts in place of a Thorlabs sensor. To facilitate proper mounting of the aperture in front of the sensor, a 3D-printed case may be utilized, enhancing practicality and accessibility for educational settings.