Recognizing Human Interaction Proofs

HIPs, Human Interaction Proofs, is used as the first gate to protect network servers from attacking, and keep users' private information safe. HIPs act like a Turing machine, when try to access to service, HIPs will provide one picture and ask the client to supply the answer. Through analyzing the response, HIPs will automatically classify the client as human being or code, and then take predefined actions. The rational is the gap between natural intelligence and machine perception.In the context of HIPs image, many factors, such as textural background, random noise, occlusion, and deformation, decrease the probability of being hacked. This thesis is dedicated to deal with deformation.First, we introduce our recognizing system and details about implementing the system.Then, we use graphical model to model the generation process of HIPs. Under the control of the inner variables, the distribution of the characters extracted from the HIPs images is one low-dimensionality non-linear manifold embedding in high-dimensionality feature space.We propose one method to model this distribution efficiently crossing over the non-linearity. In this thesis, we use piece-wise linear manifold to approach to the non-linear manifold, transferring the global non-linearity to local linearity. Through experiments, the system's performance is boosted.At the end, our model can be applied to recognize the handwritten digits. For MNIST our result is 96.54%, and for USPS the result is 96.32%.