Research And Implementation Of Image Saliency Prediction Based On Deep Neural Network

Humans have a highly developed visual attention mechanism during their long-term evolution.They have the ability to quickly and selectively discover interesting targets in a complex environment.Simulating this visual attention mechanism has attracted the attention of many researchers in the field of psychology,the nervous system,especially in the field of computer vision.Image data is the main carrier of visual information,when watching a picture,the human visual attention mechanism makes their fixation points on the picture fall in special areas which are more able to stimulate our vision,these special areas in the image are called salient area.The saliency prediction is to imitate this visual attention mechanism to extract the important information of the image and apply it to the new field of artificial intelligence.In this paper,two models of image saliency prediction are established by training and learning method.The emphasis of this paper is to apply the deep learning technique that has been popular in recent years to the saliency prediction.We combine the advantages of deep convolution neural network with traditional saliency principle to construct the deep learning model.In the first model,the high-level and low-level semantic information of the image are taken into account when training the deep neural network,and the convolution neural network with multi-scale inputs is used.In the second model,the traditional saliency detection methods based on global contrast and local contrast of a image are combined with deep learning,and the recurrent convolution neural network is used.Thus,the information of various levels contained in the image are considered.The two kinds of network models in this paper,multi-scale input convolution neural network and recurrent convolution neural network are also useful for other fields of image processing and recognition.The periodic training method proposed in this paper can achieve the goal of automatically refining the saliency map by combining the original input image and the saliency map of preliminary training.In addition,we make the experimental results of the two models more accurate by blurring the generated saliency map.In this paper,two depth neural network models are implemented with the platform of Caffe deep learning framework.We use the software of Matlab and Python to preprocess the natural images,import and export the model results.We make experiments with the two models on the MIT,SALICON and iSUN databases and evaluate them from a qualitative and quantitative point of view.The results of model one and model two are better than the saliency prediction results of traditional model and common convolution neural network model.