Research On Iteratively Learning Based Weakly Supervised Semantic Segmentation

With the continuous innovation of deep learning models and the improvement of parallel computing ability,deep CNNs-based methods have made tremendous progress in image semantic segmentation task.However,these methods require a large amount of pixel-level class labels,which requires a lot of human resources to annotate such accurate object masks.Therefore,a well-designed weakly supervised semantic segmentation method,using bounding box or image-level class categories,is an important way to alleviate the limitations.In recent years,some handcrafted methods,like MCG or Dense CRF,are often used to generate pseudo labels for training the semantic segmentation network.However,these methods do not generate accurate pseudo labels specifically for these weak labels like bounding box or image-level labels,which often limit the performance of the trained segmentation model.Therefore,there is still much room for improvement in weakly supervised semantic segmentation task.In this paper,we propose a learning-based pseudo mask generation(LPG)framework for weakly supervised semantic segmentation task with a bileveloptimization training strategy based on EM algorithm.We use auxiliary dataset with pixel-level labels to train the pseudo mask generation network iteratively,so as to map the rough segmentation results learned by weak labels into more accurate pseudo masks,and further improve the segmentation performance of the network.Once the pseudo label generation network is trained,it can be applied to the any new dataset with only bounding box or image-level class categories,for providing more accurate pseudo labels to improve the performance of the semantic segmentation networks.We have conducted experiments on both box-supervised and image-level supervised semantic segmentation tasks,the results demonstrate the effectiveness and generalization ability of our learning-based pseudo mask generation method.We achieve the state-of-the-art performance both on box-supervised and imagelevel class label supervised semantic segmentation tasks.