Object-Recognition Based On Bayesian Inference

The human vision system is the most sophisticated object recognition system, which has no difficulties in interpreting scenes containing tens or even hundreds of objects of different sizes, shapes and poses. Computer vision aims at to duplicate the effect of human vision by electronically perceiving and understanding an image, and then complete the task under the specific environment in special duty. Object recognition is currently one of the hottest direction in computer vision field, has a wide application potential in the security, business and other fields. For the particularity of the problem of object recognition, it is also one of the very difficult problems in pattern recognition field.This thesis addresses object recognition based on Bayesian inference: from filtering the image using a Gabor filter, modeling the feature appearance of the object, aligning the training shapes, modeling the shape of the object, to combining the model of the feature appearance of the object and the model of the shape of the object in a Bayesian framework, sampling the locations of the features by using Sequential Monte Carlo, finishing the establishment of the object recognition model and the training and testing of the model. Furthermore, we give the satisfactory experimental results. Based on above, we discuss as follows.1,Bayesian probability theory and sampling methodsBecause the probability distribution of samples is not well known in advance, it may be some kind samples are collected in relatively full, and another kind samples are not collected when collecting the samples, then this leads a great deviation when statistic learning, so object recognition problem is a problem of mathematical uncertainty. According to the Bayesian point of view, all observed and unobserved quantities are considered random variables, so Bayesian theory is suitable for describing and solving the problem of mathematical uncertainty. As a result, to solve the problem of object recognition by Bayesian theory is very reasonable. In Bayesian theory, the base is Bayes theorem. In some cases we do not know the prior probability of some variable, we need to assign a reasonable probability distribution to the prior distribution, making model includes different levels of parameters, namely, a hierarchical model. For Bayesian statistical model there often need the analysis of multidimensional integral, if in addition to the basic probability distributions, the model contains other forms of probability distributions, then posterior probability formula is not closed and the problem may be more difficult to solve. For this case, a popular solution is to use sampling algorithm. Currently some popular sampling algorithms are Metropolis-Hastings algorithm, Gibbs sampling algorithm and SMC, which have advantages and disadvantages, for a number of different occasions. In this thesis we use SMC.2,Gabor filter and feature appearance modelResearch shows that the response to visual stimulation of simple cells in the visual system can be approximated by using Gabor wavelet. Gabor filters are direction-sensitive edge detectors well-suited to feature matching tasks. This paper will use Gabor filters convert the image, making the features more distinctive in the image. After getting the filter response values, jets, we transform the Cartesian jets into hyperspherical angle coordinates and assign the energy of the jet to each feature, then assume that the distribution of the jets and the energy for each feature follow a Gaussian distribution independently, so we have the feature appearance probability model.3,GPA algorithm and object shape modelAn object shape can be considered as a planar configuration of the feature locations, different spatial relationships between features can constitute different object instances, and the spatial relationships between features are subject to certain limitations, such as the permissible range of the deviations of those features, the relationship between points and so on. For the training images, as there may be a certain deviation of the locations between camera and objects when the objects were imaged, there are a certain scale effects between the training shapes. To remove scale effects from the training shapes, in this paper, we use an improved GPA algorithm to align all training shapes. In order to establish the object shape model, we assume all features obey a Gaussian distribution, Because we do not have any real prior information about the covariance, in this paper, according to some literatures, we set a vague conjugate inverse-Wishart prior on the covariance matrix. With setting some reasonable values on these parameters, we get the object shape model which is approximatively a joint Gaussian distribution.4,Object recognition modelIn the model, our observed variables are the transformed image, the training Gabor jets and training object shapes; our unobserved variables are the locations of the target features and the shape model hyperparameters. In this paper, the object recognition problem can be described as estimating the posterior distribution of the feature locations given the image and the training data. After some reasonable assumptions and factorization, the posterior distribution can be described as the combination of the feature appearance model and the object shape model.After getting the object recognition model, we learn and test the model. The training process is summarized in the following:(1) Reading the training images and the feature locations of each training image.(2) Filtering the images using Gabor filters, we get the amplitude values and the phase values, namely, Gabor jets.(3) Transforming the Cartesian jets into hyperspherical angle coordinates, computing the energy at each feature location.(4) Using the values in step 3 and the feature appearance model, we get the parameter values of the feature appearance model.(5) Aligning the training shapes, we get the parameter values of the object shape model by the aligned shapes and the object shape model.(6) Using the parameter values in step(4) and step (5), we get the parameter values of the object recognition model.For a novel image, in order to get the expected locations of each feature, we use SMC sampling method. The testing process is summarized in the following:(1) Transforming the image using Gabor filters and getting the response values at each pixel.(2) According to the response values and the parameter values of the model we sample the posterior distribution and get particles and weights, then estimate the feature locations and the posterior.Finally we give the experimental results and the analysis of experimental results.