Thursday, May 23, 2019

Fast Imaging in the Dark

Pixart and National Chiao Tung University, Taiwan publish an open access paper "Fast Imaging in the Dark by using Convolutional Network" by Mian Jhong Chiu, Guo-Zhen Wang, and Jen-Hui Chuang presented at 2019 IEEE International Symposium on Circuits and Systems (ISCAS):

"While fast imaging in low-light condition is crucial for surveillance and robot applications, it is still a formidable challenge to resolve the seemingly inevitable high noise level and low photon count issues. A variety of image enhancement methods such as de-blurring and de-noising have been proposed in the past. However, limitations can still be found in these methods under extreme low-light condition. To overcome such difficulty, a learning-based image enhancement approach is proposed in this paper. In order to support the development of learning-based methodology, we collected a new low-lighting dataset (less than 0.1 lux) of raw short-exposure (6.67 ms) images, as well as the corresponding long-exposure reference images. Based on such dataset, we develop a light-weight convolutional network structure which is involved with fewer parameters and has lower computation cost compared with a regular-size network. The presented work is expected to make possible the implementation of more advanced edge devices, and their applications."


  1. I think advancements in denoising have been remarkable over the last few years. In this work, we compare Fig 1 image (c) to denoised image (d). To me the denoised image (if you enlarge it) looks like a painting so I suppose it might be an improvement for certain applications. Maybe a negative for others.

  2. This is a denoising algorithm and it can not be generalized to other application. I think another algorithm for low light enhancement which was introduced in 2018 gives better results.
    The algorithm is called Learning to see in the dark. it can be found here:
    two main problems of the See in the dark are: complexity, generalization.
    Complexity can be solved by using a smaller U-net network. For Generalization problem, it is possible to generate a large data-set of different image sensors and train the network. (or generate a synthetic dataset indifferent light conditions).

  3. I believe much of those AI algoritmhs are creative guesswork in order to resemble something, not nesecerily the truth, but something. A fuzzy zone between reality and fiction. I believe there must be users that dont mind a high degree of fiction, and others that dont want it at all or at a lesser extent. AI should be optional for the end user.

  4. Those algorithm seems to perform some kind of assisted image synthesis.


All comments are moderated to avoid spam and personal attacks.