Research On Visual Attention Based Neural Network Model And Its Application

The urban intelligent monitoring system plays an important role in the traffic and safety management of the city.The research of computer vision technology provides a solid technical foundation for the city intelligent monitoring system.Artificial neural network technology has become one of the most popular technology in the field of computer vision.In recent years,the research of artificial neural network technology mainly through increasing the complexity of network model to improve the network performance,and achieved many excellent results in the task of image classification,however,this idea has met the bottleneck.The artificial neural network can also improve the research and simulation of the working mechanism of human visual system,which has not been widely studied.In view of the above problems,this dissertation takes the simulation of human visual system as the starting point,and carries out the research of neural network model based on visual attention mechanism and its application.The main research work and innovation achievements of this dissertation are as follows:1.Proposed a neural network model based on visual attention mechanism:Sal-Mask Net.This dissertation proposes a Sal-Mask network model which can simulate visual attention mechanism.The Sal-Mask network is a network model formed by introducing saliency map as an additional information on the reference network,and the Sal-Mask connection combination which selectively enhances the feature extracted from different regions according to the mask map.Its working principle is to score different regions of the input image according to the saliency map,enhance the features extracted from the important area,and weaken the features extracted from the less important area,so that the network can concentrate on the features of the important area during the training process,and filter out the interference of the less important area on the training.The ex-perimental results show that:When using the appropriate saliency mask map,Sal-Mask network model can effectively improve the classification performance of the network.Sal-Mask network can work effectively for the benchmark networks with different data sets and different structures,and it is a universal network model.Sal-Mask network barely increases the complexity of the network,its increasement on performance of the network is not gained by increasing the complexity of the model at the cost,but by improving the efficiency of the network to achieve.In addition,we can also observe that the performance of Sal-Mask network is directly related to the performance of the saliency map used as a mask map.2.Proposed a mask algorithm based on artificial neural network training:Ada-Sal Net.Because the performance of Sal-Mask connection proposed in this dissertation is directly related to the effect of the saliency map,in order to solve the problem of ob-taining better mask map,this dissertation proposes a mask algorithm based on artificial neural network training;Ada-Sal.The Ada-Sal network consists of the main network responsible for the image classification task and the sub network responsible for generat-ing the mask map.The main network and the sub network cooperate training,feedback each other,and improve the classification performance of the network while getting better mask map.The experimental results show that:Through this kind of cooperativetraining,we can get a better map of the saliency map,and the classification performance of the network has also been significantly improved.Through summarizing and com-paring many groups of experiments,this dissertation also summarizes some guiding principles that should be paid attention to when designing a sub network for training mask map.3.Proposed a neural network model with multi-scale feature selection en-hancement.The essence of the Sal-Mask network is to selectively enhance or weaken the char-acteristics of the input image according to the saliency score of its corresponding region.Based on this model,this dissertation puts forward the Deep-Sal network which acts on the deeper layer by extending the depth.Subsequently,in the breadth of the extension,a global scale of the feature selective enhancement of MOF network,and the multi-scale feature selective enhancement of the Sal-Mof network model.The deep sal network selectively enhances the features of the deeper layers in the neural network.Each chan-nel in the feature graph extracted from the network is trained to obtain a weight value.according to this weight,the global scale of the feature map is selectively enhanced,while Sal-Mof is a multi-scale feature selection enhancement model which simultane-ously processes the features on two different scales.The experimental results show that the performance of the model can be further improved after the extension of the Ada-Sal network in depth and breadth,especially the Sal-Mof network with feature selective enhancement on multi-scale can greatly improve the performance of the network.