Computational Models Of Visual Attention

Visual attention mechanism can be used to select most important information and ignore the irrelevant or interfering information. It can also be used to select the most relevant parts of a scene according to the goals of perceiver. Thus, visual attention can guide eye movements. How to simulate visual attention and apply it to various aspects of computer vision has been a hot topic in this field.By analyzing and simulating the computational principles of visual attention mechanism, three computational models of visual attention are proposed according to bottom-up attention or top-down attention. All of the models are used to predict human fixations in complex natural scenes.First, we introduce a local contrast based bottom-up saliency model for predicting human fixations in free-viewing natural scenes. By modeling the center-surround mechanism, saliency measure is defined as the difference between statistics of center and surround. Features are fused by exploiting second-order statistics and first-order statistics of features, which are represented by region covariance and mean respectively. Taking advantage of the independence between the size of region covariance and mean and the size of image region, multi-scale saliency can be implemented directly without down-sampling and interpolation. The model is compared with 12 existing bottom-up saliency methods both quantitatively and qualitatively on three public benchmark data sets. Our model outperforms most of the models, which demonstrate our model is an effective predictor of eye movement.Second, we present a bottom-up model of visual attention based on global contrast. The saliency is computed by comparing each patch between all of other patches in the whole image. The high level features are extracted by applying independence component analysis (ICA) to each patch. The global statistical information is presented by histograms of ICA coefficients of all patches in the whole image according to each independent component. The global contrast based saliency is computed based on the histograms. Our experiments on three benchmark data sets demonstrate that the proposed approach is effective for predicting human eye fixations and is highly competitive compared with the state-of-the art saliency models.Last, a top-down computational model based on attentional mechanism is proposed for predicting of human fixations in object searching tasks in natural scenes. Three factors are integrated into the model:bottom-up saliency, target appearance, global scene context. The bottom-up saliency component based on center-surround mechanism is measured by the Incremental Coding Length (ICL). The target appearance along with the global scene factor introduces top-down knowledge into the proposed model. Our model has been validated against eye movements of human observers performing object search tasks in real world scenes and demonstrated to be effective and comparable to state-of-the-art computational models of object search.