Autonomic Learning Of Semantic Segmentation Based On Visual Attention Mechanisms

In recent years,as one of the hot research topics in computer vision,semantic segmentation has made great progress,especially after the emerging of convolutional neural networks(CNNs).Because of the large-scale datasets with tens of thousands of training images,the precision of CNN-based segmentation models rises.However,fully-supervised semantic segmentation networks heavily rely on tremendous humanlabeled annotations and hence needs significant human labors when dealing with new categories.Weakly-supervised semantic segmentation,due to its dependence on less annotation data,gradually becomes another hot research topic.Early weakly-supervised semantic segmentation methods mostly leverage tools like attention models to generate seed areas to train segmentation models.In addition,salient object detection models,due to their ability to capture foreground regions,have been being adopted by many approaches.By appropriately combining attention maps with saliency maps,these approaches can allocate each pixel in the class-agnostic saliency maps a class-specific tag,which makes training segmentation models possible.The aforementioned description is based on an assumption that the image-level labels associated with each image are precise.In fact,given a collection of category keywords,how to extract useful information from the free web images to train segmentation models is of great interests.Taking the above challenge into account,this dissertation aims at presenting an effective way to tackle this general weakly-supervised semantic segmentation problem by leveraging two different visual attention mechanisms(salient object detection and attentive region detection).Beyond that,regarding the drawbacks of existing saliency and attention models,this dissertation also introduces useful thoughts to advance them.The contributions of this dissertation can be summarized as follows:1.Presenting a salient object detection model based on top-down short connections.By connecting side paths to the stages of classification networks and building short connections between each pair of side paths,high-level semantic features can be delivered to lower side paths and low-level features with rich edge infor-mation can help refine coarse high-level features.Experiments on 5 widely-used benchmarks show that the proposed approach improves all existing methods.2.Presenting a self-erasing strategy for attention models.Based on the adversarial erasing strategy,the proposed approach takes the location of background as priors and designs two different self-erasing strategies,which are able to well solve the problem of the unstoppable expansion of the discriminative regions to the back-ground for models based on the adversarial erasing strategy.Experiments show that the resulting attention maps are with high quality and lead to high segmen-tation precision when applied to the weakly-supervised semantic segmentation task.3.Presenting how to intelligently learn semantic segmentation by extracting useful knowledge from the Internet.To deal with the large amount of label noise presenting in web images and their complex background,the concept of noise erasing network(NENet)is proposed.By learning semantic knowledge from discriminative regions by attention models,NENet is able to assign each fore-ground regions extracted by saliency models a prediction label and erase regions that are unrelated to the query keyword.