Research On Visual Information Encoding Model Based On FMRI

Visual information is the main way in which human beings perceive and understand thephysical world. Researchers have found that human brain can handle complex visual informationin an effective and robust way, suppressing the noise in it at the same time. With theseadvantages of human brain, it is important for Artificial Intelligence (AI) to study the processingtheory of visual perception and combine the computing ability of computers with human brain’sperception ability to the complex environment to develop new techniques for intelligent visualinformation brain-computer interface. As the best way to study brain’s function and observe itsactivities repeatedly, functional Magnetic Resonance Imaging (fMRI) becomes the main way tostudy the processing theory of visual perception. How to simulate the activities of visualperception based on fMRI is an essential problem for visual information brain-computerinterface and receives widespread attention. Taking the above into consideration, the study in thisthesis is meaningful in both theory and practice.In this thesis, we explore the method to analyze BOLD signal of rapid event-related fMRIexperiment, the optimization method to solve the encoding model and better criteria to constructreceptive field model, taking advantages of fMRI technique and the sparse principle of visualperception. Main research results are as follows:1. We investigate optimized method to analyze fMRI time-series in encoding model. Forextremely rapid event-related experiments, the BOLD signals evoked by adjacent trials areheavily overlapped in the time domain. Thus, idenfitying trial-specific BOLD responses isdifficult. In this study, we propose a new regularization framework called mixed L2normregularization, which improves the accuracy of HRF estimates in rapid event-related experiments.Then, this estimation method is applied with least square-separate (LS-S) regression model,heavily reducing collinearity induced by the correlation between trial-specific regressors. Bothsimulation study and rapid event-related four-category object classification experiment indicatethat this method could produce more accurate BOLD signal evoked by different stimuli.2. We propose a sparse optimization criterion to solve the visual information encodingmodel, using dictionary learning and sparse representation method, under the sparse codingprinciples of visual perception. Experiment results show that this method could solve theencoding model effectively, and the sparse encoding model is also consistent with principles ofvisual perception. This method leads to an effective sparse representation of encoding model andverifies the principle of sparse coding in visual perception.3. We put forward a receptive field construction method based on multi-scale decomposition of natural images, considering that visual perception system has long-term interaction withnatural environment. Firstly, regarding the sparse distribution of high-order statisticalcharacteristics, we decompose natural images in multi-scale using independent componentanalysis (ICA). Then, parameters of Gabor filters are determined by fitting the image basis toconstruct receptive field model. Experiment results imply that we can obtain more accuratereceptive field model using this method, which gives us a new way to optimize the receptivemodel in visual information encoding model.