AI Image Processing
Project Status: Completed
Image-to-Image Translation with Conditional Adversarial Network. Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros. In CVPR 2017
S. Li, H. Guo, W. Sun and X. Sun, "A low-illuminance image enhancement method in YUV color space," ICMTMA 2022, pp. 286-291, doi: 10.1109/ICMTMA54903.2022.00062.
Project Type: Internship
Project Timeframe: 6/1/2020 - 9/1/2020
Categories: Image processing/computer vision, coding, biomedical robotics
Skills/Software used: python, shell script
This page contains images of the GI tract, please view at your own discretion!
Anx Robotica Corps. is a biomedical engineering company that also specializes in robotics. Their star product, the capsule endoscope, is a pill-sized robot that can be controlled via an external magnet, offering a painless, non-intrusive alternative to traditional endoscopy.
​
Since the only light source inside the GI tract is the capsule itself, however, the images and videos captured by the robot often has bad quality, contrast, and definition. The ALTM image processing algorithm is a solution to this issue, as it brings up the luminance and contrast and thus making the media more readable. This algorithm takes huge amounts of computing power and time, however, and with tens of thousands of images outputted by each endoscopy session, this algorithm becomes too time-consuming to use.
​
I worked as a software/computer vision intern at Anx Robotica to use machine learning to train and AI model for ALTM, which will be significantly faster than using the algorithm itself. For training scripts and a fuller selection of training images, please visit my github page.
Training requires using python and shell scripts to generate thousands of training images and feeding them into the machine learning algorithm. Some results are shown below; real A is the original image, real B is the ALTM algorithm processed image, and fake B is the image from the trained AI model.
High Definition Image Training
While the ALTM model takes care of the contrast and luminance of images, it does not improve the definition aspect. Here, there is no advanced model needed; there are many ways to train an AI model to improve image definition. The method I chose to use was to write python scripts that decrease the definition of images using pixel, area, or cubic averaging, then using the low-def image to train towards the original image. A few examples are shown here; a fuller selection can be found on github.
