Batch Mode Active Learning For Semantic Segmentation

Image semantic segmentation refers to the technique of classifying and labeling each pixel with semantic information in the image.It is a basic research topic of the field of computer vision,which has a great significance for scene understanding and environment perception.However,large reliable and efficient labeled datasets are required for training a powerful semantic segmentation model and it is very expensive to construct pixel-wise annotated images.In order to solve the problem of lack of effective datasets and the time-consuming annotation,we apply the active learning algorithm selecting more meaningful samples to be labeled from the unlabeled dataset,and construct a reliable and effective dataset.We propose the following work in combination with the characteristics of semantic segmentation task:(1)A batch mode active learning algorithm with prior information for semantic segmentation is proposed.In order to make better use of the relationship between pixels,this paper firstly introduces the prior information of Nonlocal Self-Similarity as a selecting criterion of active learning.This active learning algorithm combines Nonlocal Self-Similarity,informativeness and representativeness three criteria and each criterion querys the subset,which is an active learning method with semantic segmentation characteristics.In addition,this paper also proposes a new measure of uncertainty for the measurement of image uncertainty.According to the selection strategy,it iteratively selects samples from the unlabeled data pool,constructs a reliable and effective dataset for training the semantic segmentation model.Aiming at the algorithm,a certain benchmark experiment was designed to verify its validity and feasibility.(2)A general batch mode active learning algorithm based on multi-clue sample selection for semantic segmentation is proposed,which combines image edge information,informativeness and representativeness three selection strategies.They can measure one sample's importance from different views and different selection strategies can achieve diversity of the training set.The three criteria select the samples composed of the selected dataset,and the labeled and unlabeled datasets are manually updated.The reliable and effective training set is constructed in the process.This method first introduces image edge information into the active learning algorithm,which can make up for the lack of informativeness and representativeness criteria and select more effective samples for semantic segmentation tasks.Aiming at the algorithm,a certain benchmark experiment was designed to verify its validity and feasibility.In general,we propose batch mode active learning methods combining the task characteristics of semantic segmentation.These methods can construct a reliable and effective semantic segmentation data set.The methods in this paper can not only reduce annotation cost,but also improve the segmentation performance for semantic segmentation.