Design And Implementation Of Saliency Detection System Based On Deep Learning And Image Segmentation

Image saliency detection is an integral part of the image analysis field.In recent years,as an important preprocessing step in the field of computer vision,it has attracted the attention of a large number of scholars.The saliency detection technology has been widely used in the fields of object tracking and recognition,image compression and detection,video retrieval and detection.From the perspective of human perception,the characteristics of an image can be represented by three different levels: low-level or intermediate and advanced features.The low-level features include the underlying information of the image,such as color,texture and so on.Intermediate features often include object information,as well as shape and spatial information.In contrast,advanced features often contain some semantic information inherent in the object.The substantial significance test is related to the three levels of features of the above image.Therefore,how to effectively integrate the above features for target saliency detection is the key to solving the problem.At present,the commonly used saliency detection methods mainly revolve around two types of top-down and bottom-up approaches.The bottom-up approach starts from a large number of low-level features,such as the color and texture of the area.It can effectively detect the details of an image,but it cannot detect global information.In contrast,the top-down approach is aimed at representative features in training samples that can detect certain fixed size and category goals.However,the top-down approach may have rough results and lack detailed information.Although many good algorithms have appeared in this field,it is still more difficult and challenging to develop an effective saliency target detection algorithm.First of all,In order to solve the huge amount of computation divided by pixel points,the preprocessing method used is the SLIC superpixel algorithm,which transforms the image from pixel level to regional level.Then feature extraction is performed on a single superpixel region.The pre-trained VGG model is used as a feature extractor to obtain deep learning features,and a similarity matrix is constructed through Euclidean distance.Then through the community detection method,the image is adaptively segmented.Secondly,Under the framework of Caffe,a top-down deep saliency detection model was obtained by fine-tuning the VGG model.The bottom-up saliency detection is performed using the community based on graph cut method,and the image segmentation method is used to refine the boundaries of the image.Then combining the two models to get the final significance model.The research content of this paper mainly includes three aspects:(1)The basic principle and application of image segmentation algorithm are introduced systematically,and an adaptive image segmentation method based on deep learning features and community detection is proposed for the huge computational problem of pixel segmentation.(2)Explained the principle of saliency detection and the typical saliency detection method.According to the deficiencies of the existing methods,an image saliency detection method based on deep learning and graph cut fusion is proposed.(3)A software system including image segmentation and image saliency detection are designed and implemented by MATLAB tool.The system can perform image segmentation on any input image to obtain a region of interest.And the input image can be divided into different regions,and significant values are calculated to obtain a saliency map.Finally,the performance of the segmentation map and saliency map are evaluated by evaluating the indicators.