Research On Image Recognition Based On Visual Saliency

With the development of science and technology, machine vision has been widelyused in many areas. In order to simulate the mechanism of human visual system, itneeds to combine multiple disciplines like computer science, biophysics andpsychology etc. Typically, when facing with a visual scene we always pay attention onthe place which has distinctive feature. According to the capabilities of giving focusof attention a priority, there have been a lot of salience based computable models. Atthe same time, many scholars have combined task-driven model with visual attentionto simplify image classification. In this paper, we also focus on object recognitionsimulating visual attention.This paper introduced the physiological basis of human visual, then detailed theimage classification method which is the integration of bottom-up saliency modelproposed by Itti and Bayesian visual attention computational model. Recognitionmethod combined saliency with Bayesian extracts saliency features, and selectGaussian model to fit the distribution of feature vectors as prior knowledge of objects.This paper proposed some improvements on this model to enhance the classificationrate.Secondly, we employed Gaussian mixture model instead of single Gaussianmodel to estimate the likelihood probability distributions of features resulting fromeach feature map. The original model chooses single Gaussian model, however, withthe increase of the training data, the feature distribution couldn’t be simply modeledas a single normal distribution. The paper proposed an adaptive method to estimatethe distribution of features automatically according to the distribution of featuresbased on mixture Gaussian model. The assumption of our method is that distributionof each feature can’t be estimated by single Gaussian function. We introduce twojudgments to make a choice between single Gaussian and mixture Gaussian.We attempted to employ a better algorithm to represent statistically uniquelocations in each feature maps, by setting the salience-based weight. We take morecount of the features with higher salience during the calculation of the likelihood. The improved algorithm is a new model of attention guidance and recognitionwhich exploits large images classification. When this model is provided severalimages from object, it will build a description of this object, and output a probabilitylist describing the likelihood that the object can be found at each input images.Experiment results on ALOI image database show that the improved method canrecognize objects more accurate and quickly.