Image Saliency Detection Based On Deep Neural Networks

With the fast development of "Internet +" intelligent information technology,huge amounts of images and video data are being stored and transmitted in the living space,which is increasingly digitally driven.There is an urgent hope that computers can help people process these data quickly and accurately,and then quickly get the important information they want.Saliency detection is to let the computer simulate the work of people or primates in this moment,which attempts to locate the most noticeable and eye-attracting object regions in images,has been widely used in such domains as vision and pattern recognition,semantic segmentation,target tracking and so on,and has shown very important research value.In this paper the main work and research results are as follows:(1)Considering that lots of saliency methods suffer from poor robust detection,this paper proposes a new saliency prediction method for global and local estimation using the combination of deep convolutional neural network and spatial transformational neural network.The candidate object areas are conformed via preprocessing methods of the removing mean and normalization.Because the saliency region is related to the semantics of the scene,a three-layer global model with convolution layer is designed to obtain the context information of the scene.On the other hand,the spatial transform network is regarded as a kind of attention technology,and combined with the convolution layer to design an additional local model for training,to learn the region of concern in the image.The output of the spatial transformer network local confidence coefficient is introduced into the global information saliency map to seek the maximum value of the feature expression.The experimental results show that the proposed algorithm improves AUC accuracy under the same condition,generates a saliency map of focus highlighting and achieves impressive robust detection results.(2)For the above method may not extract the learning features of multiple scales for saliency detection,on the basis of VGG16 network structure,this paper proposes a multi-scaled learning based saliency detection CNN model.The model compositely considers low-level and high-level semantic information of the image,and combines multiple layers of information to obtain multi-scale features.First,a CNN network structure combining bottom-up and top-down feature information is designed to determine the pixel size of image saliency region detection.Then,the improved VGG structure is copied to form three parallel network streams to handle the zoomed-out version of the same input image.Finally,a multi-scale fusion CNN network is constructed to further determine the weight of the output of each significant map in the three result graphs,and weighted by the group of weights to obtain the image salient area.It is proved that a good detection effect has been achieved.(3)According to the analysis and realization of the above algorithm and based on the PyQt5 graphical programming technology,this paper design an image salience detection software.The test results of several selected images show that the software run smoothly with simple interface and good interaction,making the saliency detection process easier and greatly improving the real-time efficient of detection.