Visual Saliency Prediction Model And Application Based On Crowdsourced Gaze Data

The visual saliency prediction model simulates the visual attention mechanism of humans through computational modeling,and predicts the degree of visual attention received by different regions in the image,which can be applied to the fields of image processing and visual design.The visual saliency prediction models based on the features of image itself have the problem of poor accuracy and scalability;while the visual saliency prediction models based on deep neural networks can effectively improve the prediction performance,but require a large amount of training data to achieve good results.The introduction of gaze data can help to improve the existing visual saliency prediction model,but the traditional eye tracking method is limited by equipment cost,space,operation process,etc.,the acquisition of gaze data is ineffective and has low practicability.To this end,this paper adopted the gaze data acquisition method based on crowdsourcing mechanism to replace the traditional eye tracking,and established a crowdsourced gaze dataset,on this basis,trained a visual saliency prediction model based on fully convolutional neural network structure,and used static webpages for testing and application.The main research work includes the following three parts:(1)Improved crowdsourcing eye tracking method based on gaze recall.Based on the original crowdsourcing eye tracking method of gaze recall and self-reports,in order to improve the accuracy and acquisition efficiency of crowdsourced gaze data,this paper studied the different parameters in the data acquisition stage and their impacts on the gaze accuracy through user experiments.This paper studied corresponding optimizations for mask pictures,task types,stimulation pictures,etc.,so that the accuracy of the crowdsourced gaze data reached one degree of visual angle,which is 3.57% higher than the existing methods.On this basis,a crowdsourcing eye-tracking dataset for static web page images was established.(2)A visual saliency prediction model based on crowdsourced gaze data was proposed.Taking static webpages as an example,a visual saliency prediction model was established using a fully convolutional neural network.The model and the heat map generated by the corresponding crowdsourced gaze data were used as input for model training.By minimizing the loss value,the model could further learn the visual perception features of the human eye.The test results showed that the Pearson's Correlation Coefficient was increased by 44.8% and the Kullback-Leibler divergence was reduced by 13.9% compared with the existing model,which verified that the prediction accuracy of the visual saliency prediction model in this paper had been effectively improved.(3)Application of visual saliency prediction model.Designed and developed a visual saliency prediction and auxiliary design system.Users could obtain visual saliency prediction results for static webpage pictures online by uploading pictures as well as corresponding design assistance.The experimental results showed that,with the help of the system in this paper,the revised design of the web designer improved the user rating by 8.18% compared with the original design,which verified the feasibility and practicability of this system.