It is no question just how ubiquitous technology has become in today’s age. We are exposed to an astounding number of digital images and video either through our beloved films, video games, and even pesky ads that pervades our social media feeds. While this is an interesting development regarding our evolution as a species, in an engineering perspective, this poses a serious data concern in terms of storage and transmission. Without a way to reduce the amount of data we use to store these images and videos, the energy and silicon used will be unsustainable in the near future as more images and videos are uploaded to the internet. As such, this project will tackle this problem and aim to minimize the data used for a given image and maximize its image quality. That is, the project will find the largest JPEG quantization parameter for a given image such that if any more compression is done onto the image, there will be a stark difference in the quality of the image.

Previous works have used metrics defined by the images quantitative properties (e.g. PSNR, SSIM) to determine the quality of an image. There are two flaws to this technique. Firstly, using this metric requires the compressed image to be compared with the the original (or pristine) image. Not only does this technique stores two copies of the same image of varying quality, but it most cases, the pristine image does not exist as many images in the internet go through an innummerable amount of transformations (e.g. other compression passes, downscaling, etc.). Therefore, we do not solve the data issue, and these metrics cannot give us an accurate conclusion regarding the compressed image’s quality. Secondly, metrics that compute a result based on quantatitive information only approximate the quality perception of real human beings. What this project aims to do is capture how humans perceive image quality and compress an image accordingly.

As the project requires a dataset of human decisions, the main method of data collection will be through Amazon’s Mechanical Turk (MTurk), which is a crowdsourcing platform that involves the distribution of a virtual task to a large human workforce. The task will allow workers to control the quality of the images through a real-time slider that compresses the image according to the slider value so that they can determine the perceived optimal point of compression. The collected dataset will be fed into a neural network that we will construct, customize, and train so that the model can learn and eventually automatically predict the optimal quantization parameter for any input image.

This project’s tech-stack:

• Amazon MTurk, Amazon S3, and Amazon AWS
• JavaScript for creating the MTurk tasks
• PyTorch for creating and training transfer learning image models
• High-Performance Computing for training the models

To read more regarding this project, I invite you to click on the PDF icon or the presentation icon on the top right of the page.