On Google's popular video sharing site YouTube, one of the finishing touches a user completes before publishing videos onto the site is choosing a thumbnail. These still frames selected to represent a video need to inspire potential viewers to click on the link and watch the creator's content. Unfortunately, YouTube seems to be plagued by poor quality thumbnails, and not only does it make finding quality content more difficult than it could be, it also causes some videos to be passed over when they would otherwise be well-received.
YouTube gives uploaders the option to create a thumbnail for their videos or hand over control to the site and pick from options given to them. Most high-profile YouTubers will generally upload a custom made image, but casual users choose from the site's suggestions. In the past, these thumbnails have been less than stellar. Luckily, the Google research team has been busy updating the service and has found a way to improve the images YouTube will present to its content creators. The team is hoping the improvements will encourage viewers to click on videos that they would otherwise not be interested in. The advancements work through a type of deep neural network, abbreviated DNN. Developments in the way computers are able to "see" and categorize images and videos are allowing YouTube to update the technology behind its thumbnail submissions. The new system work by taking a frame from the uploading video for each second of its length. Next, the chosen frames are given individual quality scores. The images with the highest ratings are offered special treatment in the form of enhancements before being made into thumbnails.
The quality scores were difficult to design, which is where DNNs come in. The team first trained its quality checking DNN so that the network would know what to look for in a contender for a video's thumbnail. This feat was accomplished by presenting thumbnails to the DNN the team deemed high-quality. These were used as positive examples. After that, the team found low-quality examples from around YouTube and gave them to their DNN as negative examples, and eventually, the neural network was able to find high-quality thumbnails from a video offered to it. The research team compared the new images to ones created by the DNN's predecessor and found that the network's images were preferred over 65% of the time. YouTube has already launched the new system on its site, adding to a set of changes rumored to go live by the end of the month. If you upload a video to Google's site, expect to see newly improved thumbnails waiting for you.