Saliency Prediction Based On Lightweight Attention Mechanism

In observing a scene,one focuses on distinctive parts of the scene instead of processing all the information in the entire scene.These regions of interest to the human eye are called Saliency regions.The human brain extracts high-level information from saliency regions to rapidly understand the entire scene.Methods for simulating the selective saliency regions of the human visual system are commonly called visual saliency prediction.In computer vision,visual saliency prediction can be used to understand and simplify complex visual problems.Therefore,saliency prediction has a wide range of applications in high-level computer vision tasks,such as object recognition,video compression,video understanding,and action recognition.In recent years,due to the multi-level and multi-scale characteristics of deep learning,it is possible to accurately capture saliency regions without using any prior knowledge.Saliency prediction model based on convolutional neural networks has been used by more and more researchers use.The method used by researchers has become the mainstream method of current research.As a key component of a convolutional neural network,max pooling and stride of convolutional layer will reduce the receptive field of the feature map and cause the loss of spatial information hierarchy,affect the accuracy of pixel-level classification,and reduce saliency predictive network performance.The network used in this thesis is the SAM network.The network is composed of the VGG16 network with dilated convolution as the basic network,as the long and short time sequence memory of the additional convolution layer and the loss function.According to the existing problems of the saliency prediction model based on convolutional neural networks,the main contributions and innovations in achievements are as follows:(1)In order to solve the problem that the accuracy of the saliency prediction model based on convolutional neural network is not high due to the limited receptive field,a novel saliency prediction method is proposed from the perspective of attention mechanism,which integrates different proportions of context information by adding feature pyramid attention mechanism.So that the high-level feature map can generate better pixel-level attention and thus obtain better pixel-level classification results.Later,the convolution block attention module was embedded in the U-shaped structure of the FPA,which improved the pixel space positioning accuracy and network noise immunity for the module.(2)Because the pyramid attention mechanism uses large kernel convolution operations,Computing cost and runtime is greatly increased.In order to reduce the negative impact of large-scale convolution kernels on the network operation speed and the amount of model parameters,Based on the existing lightweight structure,a lightweight convolution operation suitable for the network structure of FPA is designed.And apply it to the improved feature pyramid attention module.Then the number of model parameters is greatly reduced without reducing the accuracy to improve the network running speed.Using this method,on the premise of not reducing the network performance as much as possible,the model size is compressed as much as possible to improve the test speed.The saliency prediction network based on the lightweight attention mechanism tested SALICON,MIT1003 and CAT2000,three common saliency data sets,and achieved good results under the saliency model evaluation parameters.