BgSub - AI tool that automatically removes backgrounds without uploading images.
Q: Why did you develop this website?
A: Because I don't want to upload my images to the web, no matter how much the site promises to protect my privacy.
Remove or replace image background without upload image - BgSub
Development Background
With the continuous development of AI, more and more AI tools have emerged, especially in the field of image processing. However, due to the huge demand for computing power in deep learning, the vast majority of tools need to transfer images to cloud servers for processing, a process that can easily lead to privacy abuse. There have been many cases of AI data misuse and privacy breaches. In particular, image processing often involves face data, which can pose a huge security risk.
The most effective and radical solution is edge computing: no data is uploaded, and the data is processed locally by the user. For deep learning edge computing, it used to require a user terminal with a good graphics card and a huge installation package to be downloaded. However, thanks to the continuous performance optimization by modern browser developers and the continuous efforts of deep learning developers to work on shrinking models, it is now possible to run deep learning models directly on the browser.
BgSub initially chose DeepLabV3 as the segmentation model and developed its own inference engine. Although the goal of in-browser computing without data upload was achieved, the whole web page reached over 300MB and the computing time exceeded 3 minutes, which was basically impractical, after all, not many people were willing to wait for a 300Mb web page to load and then wait for a long computing time. In the following eight months, the model was continuously modified, pruned and compressed, and the computing engine was continuously optimized. Finally, the whole web page was optimized to 10M and the computing time was optimized to about 5 seconds. The usability was extremely high, and the processing time was even faster than some websites using cloud servers.
Comments
Post a Comment