Research On User Interactive Image Segmentation Via Weakly Supervised Learning

Image segmentation is a technique of computer vision with high practical value,which provides mid-level vision features and is widely used in medical imaging,traffic control,object detection,etc.Recently,with the vigorous development of deep learning,image segmentation based on supervised learning has made great progress.However,supervised approaches rely on a large amount of annotated data,which limits their applications.In contrast,unsupervised approaches are annotation-free,but their performance is often far from supervised learning methods.To win the trade-off between unsupervised approaches and supervised deep approaches,we study the user interactive image segmentation methods using weaklysupervised learning.The study involves two specific image segmentation tasks,that is,texture image segmentation and single-image moving object segmentation.By using user interaction,a small amount of noisy data is obtained as weakly-supervised information,and a weaklysupervised learning segmentation model is constructed to achieve effective segmentation while greatly reducing the dependence on annotated data.Texture image segmentation is about dividing a texture-dominant image into multiple homogeneous texture regions.Since texture image segmentation is often related to specific tasks,in different tasks even the same image is often partitioned in different manners.It is difficult to adapt the unsupervised texture segmentation to specific tasks.Meanwhile,for some specific texture segmentation tasks,an effective large-scale texture segmentation database is often unavailable.To address these issues,the proposed interactive weakly-supervised approach requires the user to mark one pixel in each texture region,whose label is directly propagated to its neighbor region.Such labeled data are of very small amount and even partially erroneous.To effectively exploit such weakly-labeled data,we construct a weakly-supervised sparse coding model that jointly conducts feature learning and segmentation.In addition,the geometric constraints are developed for the model to exploit the geometric prior on the local connectivity of region boundaries.Single-image moving object segmentation aims to segment the moving objects in the image.Compared with the multiple frames used in video-based moving object segmentation,a single image lacks the temporal information,such as the relationship among video frames.Moreover,the moving objects in a single image are often obviously blurred,and the segmentation requires a clear version of the object,which greatly increases the difficulty of segmentation.To this end,we adopt a user interactive strategy,which allows users to draw a rectangular box on the image to roughly keep the moving object inside for providing some weakly-supervised information as a guide.Based on this guide information,we propose a weakly-supervised single-image moving object segmentation model.We introduce DIP(Deep Image Prior)to the layer of the moving object,and use convolution process to model the object motion.In addition,we use soft segmentation masks to model the composition of the object layer and the background layer.By solving the model,the object layer,background layer,and the corresponding soft segmentation masks can be all generated.Through experiments,we have validated the effectiveness of the proposed approaches on texture image segmentation and image moving object segmentation.The related results can provide new insights for other weakly-supervised and user interactive methods.