Answers are grounded in this paper and cite their sources.
The demand for high-quality, lightweight infrared imaging systems is rapidly increasing. Single-lens computational imaging, combining single-lens with post-processing algorithms, offers a promising solution to miniaturize imaging systems while maintaining performance. However, these post-processing algorithms are typically highly complex, posing significant challenges for real-time reconstruction on a neural network processing unit (NPU) chip. This study investigates the relationship between the complexity of post-processing algorithms and lens modulation transfer function (MTF), demonstrating that a single-lens system with highly consistent MTF can significantly reduce the complexity of post-processing algorithms. Building on this insight, we proposed an enhancing lens MTF consistency-based single-lens design method and developed a single-lens computational infrared imaging system featuring a small Res-Unet (S-Res-Unet) neural network. Compared to the traditional method using the large Res-Unet (L-Res-Unet) to achieve comparable reconstruction performance, the proposed system achieved a 16-fold reduction in computational demands. As a result, it performed real-time reconstructions at the rate of 25 frames per second (fps) with a resolution of 640×480 on the RK3588 NPU chip, while maintaining system MTF exceeding 0.42 at Nyquist frequency of 42 lp/mm. This research paves the way for the practical application of computational infrared imaging systems.
No takes yet. Share an insight, caveat, or question.
Xing et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: