3 ms·
The article just skims over this part: "To start off with, Goodfellow and co place some limits on the task at hand to keep it as simple as possible. For exampl
by softdev12 12y ago
The article just skims over this part:
"To start off with, Goodfellow and co place some limits on the task at hand to keep it as simple as possible. For example, they assume that the building number has already been spotted and the image cropped so that the number is at least one-third the width of the resulting frame. They also assume that the number is no more than five digits long, a reasonable assumption in most parts of the world."
This seems like a huge task. Someone has to go through all the thousands of images and first crop them? During that time, it would seem like they could just input the number into a database.
Maybe I'm missing something, but I read the "cracked" part to be a totally automated system that scans all the pictures and pulls the numbers with no human manipulation.
- sanxiyn 12y agoOf course cropping is also automated, but using the different algorithm. Text detection and text recognition is a different problem. Text detection is usually solved by stroke width transform. The article focuses on text recognition using the neural network.