It’s feature recognition of pixel encoding. Pixels make different rectangular Tetris shapes called HAR-like features. These shapes and features have different visual filters attached them. Similar to filters people put on Instagram filters. Eventually patterns are recognized from these Tetris shapes. If enough of these shapes are recognized by training a recursive neural network, eventually finding enough features will be considered detection. These captchas take pictures with not enough HAR-like features to allow a computer to perform object recognition. A human will still be able to perform a classification.
Latest Answers