Studies On Saliency Object Detection Based On Deep Learning

The saliency detection technique is mainly used to find areas in the image that attract human's visual attention and cognitive system,and it has an important role in the field of computer vision.And in recent years,saliency detection has attracted more and more attention from researchers.Saliency detection techniques are often used as pre-processing steps in other image processing tasks,such as image segmentation,image classification and compression,etc.,which can effectively focus on the image region related to the current task and widely promote the efficiency of the computer.Existing image saliency detection techniques based on traditional methods have made some progress,but they often use hand-crafted features or construct various prior-knowledge to calculate saliency value.When the scene in the image is more complex,the traditional saliency detection method cannot detect the saliency object correctly.Nowadays,deep learning technology is more and more mature,it has a wide range of applications in image processing and natural language processing,which has brought great convenience to people's life.Deep learning automatically combines the low-level features of the image to obtain high-level features that are useful in image processing tasks,so it has achieved good results in tasks such as image classification and object detection.The dissertation proposes a deep learning saliency detection model based on global prior and local context.Firstly,the color image and the depth map are segmented,then the global-prior based feature map corresponding to each superpixel is constructed separately,and the global prior saliency value corresponding to each superpixel is calculated by the global prior deep learning model.Secondly,the color and depth local context of each superpixel are obtained respectively,and we calculate the initial salient value of each superpixel by using a deep learning network model combined with global prior and local context.Finally,the initial saliency map is optimized based on spatial consistency and appearance similarity,and the final saliency detection result is obtained.Experiments on two commonly used datasets show that the method presented in this dissertation is accurate and has good robustness.The main work of this dissertation is as follows:1)The method for generating a global-prior based feature map is given.The dissertation makes full use of low-level features such as color,texture and depth to build the compactness feature map,the uniqueness feature map and the background feature map,then they are combined into a global-prior based feature map as a single channel respectively.2)An image saliency detection method guided by global prior is presented.Through the detailed analysis of the deep learning saliency detection method using the original image as input,it is found that the input of the detection method has invalid feature information,so we use the global-prior based feature map as an input to effectively alleviate the occurrence of the problem,and the experiments can prove the effectiveness of this measure.3)An image saliency detection method combining global prior and local context is given.The saliency detection method consists of two parts of a convolutional neural network.First,the color local context and the global-prior based feature map are taken as the input of the first convolutional neural network,and we can get a saliency map,it and deep local context are taken as the input of the second to calculate initial saliency value.Finally,an undirected graph is constructed based on the segmented superpixels,and we constructs the optimization model by using spatial consistency and appearance similarity and optimize the initial saliency map to obtain the final saliency map.Experiments show that the detection method of this dissertation has advantages over existing deep learning methods.4)We design and implement an image saliency detection system based on global prior and local context on the above work.This dissertation describes the main design and related functions of the system in detail,and analyzes and evaluates the testing results of the system.