Multi-task Weakly Supervised Learning For Image Dense Prediction

Nowadays computing resources are becoming rich and cheap,making researchers in the area of computer vision and image processing prefer machine learning,especially deep learning models.That is,to train a large model on a big dataset,which results in astonishing performance.However,these training data is manually annotated by human.Although computing resources are cheap today,human annotations are still expensive,and sometimes difficult to obtain.This makes it an interesting and important research problem how to efficiently use training data.Recently,some methods are proposed that make models learn from annotations that are incomplete but easier to obtain,i.e.weakly supervised learning.However,existing research mainly focuses on training a model on single dataset to solve one problem.Multi-task learning using weak supervision remains unexplored.This thesis works on multi-task weak supervised learning in pixel-level tasks of image processing,specifically,focusing on the following three problems:Learning to detect salient objects with multiple weak supervision.Joint learning of saliency detection and weakly supervised semantic segmentation.Iterative object removal with confidence feedback.