Session Index

Instead of traditional image models, which accompany various prior conditions and feature engineering, in recent years, data-driven deep learning network models have gradually become attractive. Based on a lensless configuration with ordinary white light (completely noncoherent), a coded mask (or diffuser) as an example to modulate the light field, the CMOS sensor measures the intermediate image, we employed the convolutional neural network (CNN) architecture to reconstruct the readily uninterpretable signal. In addition to the reconstruction, we also conducted the face recognition as the vision task. Practical implications through such unconventional optics will be expected if pattern recognition, rather than relying mostly on human recognizable fidelity as it does in conventional imaging system, become more directly linked to the hidden or incomplete optical fields. More possibilities will be addressed during the conference presentation

  Preview abstract

Real-time fall detection via deep learning is usually a challenge on embedded system due to limited computing power. In this work, we apply attention transfer to reduce the computational complexity and optimize memory usage. The proposed method achieved 98.06% testing accuracy on a real-time multi-view fall detection system.

  Preview abstract

We proposed a spatial-modulate method to recover missing data in depth images. Depth cameras often fail to get depth data due to shinny and absorbing surfaces. Additionally, depth image inpainting using deep learning methods have compromised abilities to noise. These advance techniques often fail to provide correct depth data due to the influence of noise. To address both issues, we iteratively applied several low-pass filters to recover the missing data. We evaluated our method using depth images which were manually added large noise and defected area. The results indicate the high accuracy, precision, and anti-noise ability of our approach.

  Preview abstract