An Interactive Image Segmentation Algorithm Using Attention Optimization

Interactive image segmentation refers to a technology that extracts the objects and areas of interest to users in the image by using the prior information provided by user interaction,which has a broad application prospect.The quality and efficiency of image segmentation results have a decisive impact on the quality of subsequent image processing.We propose a new interactive image segmentation model based on the investigation of early methods.The model uses high-resolution network as multi-scale feature extraction module,and integrates context of features within the possible target area.In some cases,the area of annotation from user can be predicted incorrectly,which results in decline of user experience and algorithm performance.We proposed an attention optimization method using back propagation of error term to deal with this situation.By modifying attention of high-level features in the network,the optimization method correct segmentation results according to user's annotation without a huge amount of calculation of gradient through the whole model.Experimental results show that our algorithm outperforms the conventional algorithms on four challenging datasets.Furthermore,we compare the consuming time with other optimization method,which proves significant advantages in running speed.