Computer Vision News - January 2018

• Features - Unsupervised dictionary learning and sparse coding are used to extract contours from a training set of the preprocessed images produced by step 1. Both Intermediate-level and Top-level feature learning takes place during the Features step. • Lateral connections - Pooling layers are learned from the input contours independently. The learning process then uses forward and backward message-passing in the network to identify objects. It finds a complete MAP approximate solution by solving the scene-parsing with forward and backward pass. The upper bound on the log-probability of the top-level nodes is given by the forward pass, while the backward pass calculates the approximate MAP based on the forward pass’ high scores. An additional outcome of the backward pass is the rejection of many of forward pass’ false object hypotheses. Results: For a CAPTCHA to be considered broken, it’s enough that it can be automatically solved at a rate above 1%. The authors’ RCN method effectively broke a variety of different text-based CAPTCHAs using small training datasets and no CAPTCHA- specific heuristics (figure below), at an accuracy rate far exceeding this threshold. It achieved a character-level accuracy of 94.3% in solving reCAPTCHAs (which translated to a 66.6% accuracy rate for solving entire reCAPTCHAs), 64.4% for BotDetect, 57.4% for Yahoo and 57.1% for PayPal, rates incomparable to the 1% at which CAPTCHAs are considered ineffective. 8 Computer Vision News Research Research