Research On Saliency Detection Based On Visual Perception Logic And Deep Learning

With the development of artificial intelligence technology,the application of salient object detection in computer vision has become more and more extensive,and gradually has become a very challenging task.According to the requirements of the task,the algorithm needs to detect the most salient object in an image.It is very helpful to get the interpretability of an image.So it can be used in many image processing tasks.Salient object detection is divided into traditional methods and deep learning.Traditional algorithms are difficult to extract the deep semantic features of an image,so the detection effect in complex images is poor;the method based on deep learning uses neural networks,and the training process of the network is driven by data but it lacks of corresponding prior information which results in some error and affects the detection effect.In response to these problems,this paper designs a multi-stage salient object detection network based on the feature comparison module and the human visual perception logic,that proves the effectiveness and practicability of the algorithm on multiple datasets.The main innovations of this thesis are as follows:(1)A salient feature extraction module based on feature comparison is proposed.First extract the environmental information around each area of an image,and then use the feature comparison method to interactively calculate the feature information and environmental information of the area to obtain a mask with the same size as the original feature map,and finally integrate the mask and the original feature map to enhance the significant area and suppress the non-significant area to obtain the salient feature.Also,we combined our module with R3 Net,and verified on the saliency data set that the module of this article can improve the saliency feature extraction ability of the network backbone stage,assist the network to detect the saliency target more accurately,prove the versatility and superiority of our module through the comparative experiments.(2)A salient object detection network based on perception logic and feature comparison is proposed.The human visual perception logic is to obtain candidate features through a simple pre-attention process first,and then obtain the features we need through the attention concentration process.By analyzing the characteristics of visual perception logic,we designed a multi-stage salient object network.The network first extracts the general features of an image through the general feature extraction module,then corrects the features through the feature preprocessing module,and then the special feature extraction module extracts the special features that meet the task requirements,and finally predicts the saliency of image based on the obtained special task features.This article takes the existing five publicly available saliency natural data sets as examples,tests the indicators and effects of the network in this article,and compares them with the current SOTA network.The results prove the indicators and detection effects of our network is better than the existing network.(3)A salient object detection network training method based on visual perception mechanism is proposed.In order to fully improve the performance of the network,we propose a special network training method by analyzing the human visual perception mechanism.In the training phase of the network,the backbone of the network uses the pretrained model and does not update the parameters.First the network extracts the general visual features and then uses preprocessing module to adapt the network to the current data set.This method can prevent the backbone from degrading after fine-tuning.The comparative experiment proves that this method can improve the speed of network training and the accuracy of network detection.