Research On Visual Saliency Detection Method And Its Applications

With the development of multimedia information techniques,the number of images acquired and shared by people has increased dramatically.It becomes an important and urgent problem in the computer vision field to let computer analyze and understand the content of images automatically.Saliency detection,as a key technique of image content analysis,can quickly acquire the most interesting areas in images and provide concise and effective content information for computers,which is a key step to solve many visual tasks such as image classification,retrieval and compression.With the development of this years,many excellent saliency detection algorithms have been proposed.However,in the real world,the huge number of images usually have diverse content and complex structures,and there are specific scenarios in practical application.It is significantly chanllenging to generate saliency maps accurately and solve practical issues by using saliency detection.In this context,based on saliency detection method and its application,this dissertation studies the rapid generation of object priori,background-based saliency map generation method,and saliency map selection method based on no-reference quality accessment network in post-processing stage,and the application of saliency detection in advertising video classification and advertising imgages evaluation.The study content and main contributions are summarized as follows.1.To reduce computational complexity of object priori in preprocessing stage,a fast object proposal generation method based on edge direction statistics is proposed.From the point of general object attributes and computational efficiency,the feasibility of using edge direction to discriminate objects is firstly analyzed,and an object feature based on edge direction statistics is constructed.A fast evaluation method based on information entropy is proposed to evaluate the probability of an object proposal containing objects.This method can locate the object region quickly,and provide a fast and effective prior information for saliency detection method.2.To avoid noise interference in background extraction,a background seeds extraction method based on shortest path is firstly proposed,which can obtain a number of reliable background seeds.Then,based on graph model theory,the flow ranking algorithm is used to generate background map.Finally,a fusion method is proposed to utilize the multiple groups of seeds,which improves the consistency and completeness of the saliency regions,and effectively suppresses the background regions in the saliency map.3.Since the saliency detection algorithm may generate some bad saliency maps,a saliency map selection method based on no-reference quality evaluation network is proposed.By selecting the best one from multiple saliency maps,this method can refine the bad maps and improve the performance of saliency detection.During the optimal map selection,a multi-information fusion saliency quality assessment network is designed.Due to utilize multi information,the proposed saliency quality assessment network can predict the saliency map quality more accurately.This method effectively utilizes various existing saliency detection methods and greatly improves the performance of saliency detection.4.To meet the requirement of effective management of advertisement video by video classification technology in e-commerce,the characteristics of advertisement video are analyzed,and a video segmentation method based on fragments and a deep network with spatial visual attention are proposed for video advertisement classification tasks.The former samples video frames from different segments to reduce the interference of uncorrelated frames;the latter uses visual attention model to extract the features of regions related to video categories,which effectively avoids the interference of uncorrelated regions and extracts more discriminant features.At the same time,a multi-branch deep convolution network was designed to effectively achieve the integrity of video feature expression.This method greatly improves the accuracy of advertising video classification,and expresses the feasibility and validity of saliency detection in practical application.5.To fullfill user's demand for interesting advertisement query and recommendation,this dissertation studies the advertisement assessment method based on visual attention.Firstly,this dissertation analyses the shortcoming of the evaluation criteria of the existing advertising assessment methods,this dissertation proposes a local cue by visual attention to reflect the users' attention towards product.Then,a product-based deep prediction network is designed to predict the users' attention level towards the product,and a advertising image database is constructed.The validation in the constructed database shows that The proposed method can achieve superior performance,which reflects the effectiveness of the method.