Research On Saliency Detection Of Image Based On Deep Learning

Saliency detection refers to detecting the attention regions of human eyes in an image accurately and quickly through a complete set of saliency detection models established by computer simulation of the human eye's visual attention mechanism.Saliency detection of the image has become a hot research field in computer vision in recent years.It is generally believed that the human visual attention mechanism includes both bottom-up and top-down types.A complete saliency detection model needs to fully simulate two visual attention mechanisms.However,our survey found that most of the existing models only simulate one of the human visual attention mechanisms.The detected saliency area has a certain gap with the actual human eye's attention regions,and the accuracy of the detection needs to be further improved.In addition,the inherent limitations of the existing model,including modeling methods and calculation methods,make these models difficult to adapt to all kinds of large-scale complex images with high resolution,and usually lead to the serious influence on the detection speed.Therefore,this paper presents a saliency detection model based on deep learning to fully simulate two visual attention mechanisms.The model has significant advantages in detection accuracy and detection speed,can detect the saliency region of the image accurately and quickly,and can be applied to various high resolution and large-scale complex images with a wide range of applicability.In summary,the main contents of this paper include:(1)Two new feature extraction methods for image saliency detection are proposed.In order to overcome the shortcomings of the existing methods in extracting low-level contrast features,including complexity and limitation,a low-level contrast feature extraction method using specially trained SAE(sparse self-encoder),which treats fixed size pixel units as the minimum processing granularity,is proposed.The feasibility and superiority of the method are verified by experiments.In addition,aiming at the shortcomings of the existing model such as the limited application scope of the extracting method of advanced semantic features and inadequate simulation of the top-down visual attention mechanism,a method of extracting advanced semantic features by using CNN(convolutional neural network),which is specially designed and specially trained,is proposed according to the idea of image classification.The wide applicability of the method is verified by experiments.(2)Combining the two new methods for feature extraction,an image saliency region detection model(SCS),which fully implements two visual attention mechanisms is proposed and trained after selecting the best classifier by experiment.The model extracts the low-level contrast features and high-level semantic features of each pixel unit by using the new methods for feature extraction,and classifies the pixel units by using ensemble classifier after combining the two types of features to determine the saliency and obtain the original saliency map.Finally,the original saliency map will be zoomed back to the same size as the original image to get the final saliency map.By comparing with several existing advanced detection models on three commonly used public data sets,the advantages of the SCS model in detecting the accuracy and detecting large-scale complex images are proved.(3)At the end of this paper,the SCS model is optimized by trying to find the best range of low-level contrast feature extraction and adding intermediate features for saliency detection to improve the accuracy of classification detection.Through comparative experiments,the best low-level contrast feature extraction range is determined and the SCS model is optimized by adding intermediate features.