Research On Image Semantic Modeling Based On Weakly Supervised Learning

With the development of Internet technology and portable mobile devices,the number of images and videos on the Internet is growing explosively.For example,Facebook added an average of 350 million pictures a day in 2019 according to Facebook statistics.Facing massive-scale data,how to effectively process new image data becomes an urgent problem.Image semantic modeling make computers understand the semantic content of images correctly,such as people,vehicles and so on,which provides effective solutions for image classification,recognition and retrieval.Early image semantic modeling methods only focused on low-level features,which included local features and global features.However,low-level based algorithms were limited by “semantic gap”,which cannot effectively reflect image understanding of people.With the development of deep learning,high-level visual features-based image semantic modeling algorithms have achieved impressive performance.However,since pixel-level semantic labeling requires many human resources,it cannot be well promoted in practical applications.Thus,image semantic modeling methods based on image-level semantic labels for weakly supervised learning attracted researchers' wide attention.This paper mainly studies the image semantic modeling technology based on weakly supervised learning,and combines with image saliency analysis and human visual perception,to achieve the most accurate training of deep models of image data to meet the various applications of image.The main work of this paper is listed as follows:1.This paper analyzes the advantages and disadvantages of existing image semantic modeling algorithms.In order to reduce the workload of labeling data manually,this paper proposes a manifold learning algorithm to automatically transfer image-level semantic labels to pixel-level area of the image.The algorithm does not need to rely on external detectors and prior knowledge of the data set.2.Since the potential noise of image-level semantic tags(such as tagging error or not tagging),this this paper proposes a generative model to build graphlets and object-graph descriptors to find the graphlet category,and uses Bayesian classification to label the unknown graphlet category,and removes the abnormal tags in the same known graphlet category,so as to enhance the noise tolerance of image semantic modeling.3.In this paper,image classification experiments are carried out on the VOC2012 and MSRC datasets by using the constructed image semantic model,which verifies the advantages of image semantic modeling based on weakly-supervised learning and the effectiveness of image semantic modeling based on weakly-supervised learning in noisy environment.