Weak-supervised Semantic Image Segmentation Based On Keyword Lable

With the improvement of computer computing power,Deep Convolutional Neural Networks(DCNN)have made great breakthrough progress in the field of semantic image segmentation,and have been widely used in various scenarios.For semantic image segmentation,its dense pixel prediction requires pixel-level label which is extremely costly,this limitation become a bottleneck that restricts the application of image semantic segmentation to actual scenes.As a result,the study on how to use non-pixel-level label for image semantic segmentation has become one of the research hotspots in the field of semantic segmentation.Therefore,this paper based on keyword label studies the weakly supervised semantic image segmentation.It is expected that keyword label can be converted into more accurate pixel-level label to improve the performance of image semantic segmentation models.Aiming at the problem that how to transform the weak keyword label to coarse pixel-level label,this paper uses class activation map to locate foreground objects,and proposes super pixel extension and saliency object extension methods to extend the localized seed label so that get more dense and accurate label.The super pixel extension method exploits color,spatial position and other information contained in the image itself,to produce the super pixel segmentation result.Then,uses the result to locally extend the sparse located seed label.Experimental results show that super pixel extension can effectively improve the performance of the segmentation model.Aiming at the problems of limited extension ability and inaccurate edge label of super pixel extension method,we use the saliency object detection results to further fine-tune the super pixel extended labels.The saliency object extension method combines a non-semantic saliency object with sparse super pixel labels to obtain a more complete foreground object and accurate edge information,which greatly promotes the accuracy of the pixel-level labels of the training samples.So as to better assist the training of semantic segmentation network and improve the performance of the segmentation model.This paper proves the effectiveness of the proposed algorithm and the necessity of each component through comparative experiments and ablation experiments.The result shows that our algorithm has a excellent performance for keyword supervised semantic image segmentation,especially for single target segmentation,which is more accurate and detailed.Moreover,the performance of our algorithm is obviously better than other similar research methods.