Research And Implementation Of Object Visual Attention Estimation Method Based On Progressive Learning

The human visual system has the characteristics of selective attention.How to understand and simulate the attention mechanism of the human visual system has received extensive attention from academia and industry,and a large number of visual attention models have been proposed;these models mainly understand and simulate visual scenes from the perspective of the scene observer.However,in areas such as intelligent monitoring and human-computer interaction,the development space often comes from a deep understanding of the visual attention of objects in the scene.For example,recognizing the attention of target individuals in surveillance scenarios helps to accurately predict their subsequent behaviors,etc.Therefore,modeling visual attention needs to be carried out from both the subject perspective of scene observers and the object perspective of scene participants,but the research and development of the latter is significantly insufficient.From the perspective of image pattern recognition,object visual attention is the ability of an object to focus on a specific range or specific object.Although some progress has been made in the estimation of object visual attention based on deep learning,it mainly maps the pixel space directly to the feature space in an end-to-end manner and only stays at the static estimation level.However,object visual attention is an interactive perception process between an object and a dynamic environment,which has complexity,uncertainty,and ambiguity.The feature information that can be obtained by static estimation is limited,and dynamic process information needs to be used to make up for the lack of end-to-end direct mapping.To this end,this thesis draws on the progressive characteristics of human visual attention and uses the cognitive mechanism of attention selection through detail processing and environment perception to construct an object visual attention estimation method based on progressive learning.The main research contents are as follows:1.In view of the fact that there are few public datasets for current object visual attention estimation,and there are a large number of problems such as unclear faces of objects,inaccurate head positions,and inaccurate visual attention focus positions,this thesis constructs a new dataset Auto Gaze and finishes labeling all samples.The quality,diversity,and annotation accuracy of the samples outperform existing public datasets.2.An object visual attention estimation method based on progressive learning is proposed.Combining the proposed hierarchical self-attention module with a residual network improves the geometric accuracy of gaze directions in the object visual attention estimation task;in addition,combining an adaptive multi-scale enhancement module with a high-resolution network provides better learning of space contextual features.Experimental results on different datasets show that the proposed method has superior accuracy and robustness in visual attention estimation.3.In order to verify the effectiveness of the model,method,and framework of this thesis at the application level,a data processing module,a head detection module,and an object visual attention analysis module were introduced,and a prototype system for object visual attention analysis was developed through system integration.The test results show that the system can accurately identify and analyze the visual attention of different objects and generate corresponding analysis results,which has practical value;at the same time,it also verifies the effectiveness of the technical scheme in this thesis.