Weakly Supervised Image Semantic Segmentation Based On Attention Mechanism

Deep learning,as an important branch in the field of machine learning,has developed rapidly in recent years.Image semantic segmentation based on supervised learning is one of the main ways for machines to understanding the scene through computer vision.However,semantic segmentation model based on supervised learning requires a lot of pixel-by-pixel annotations,which makes it take great time and manpower costs.In contrast,the related method based on weak supervised learning only requires image-level or object box level annotations,the lower cost of the annotations makes weakly supervised semantic scene understanding methods get more and more attention.Although it is easier to obtain image-level classification annotations or object box annotations,it is also more difficult to achieve scene understanding based on such weak annotations.The main challenge is that due to the lack of location information annotations,That is,the lack of accurate location information and boundary information of the object makes it difficult for the algorithm to distinguish the foreground object from background.In order to overcome this challenge,this paper attempts to introduce an attention mechanism into a multi-class neural network based on image-level annotations,which is used to capture object position information and boundary information to generate object pixel-by-pixel label,and to train image semantics Segmentation model.Specifically,The work mainly includes the following two aspects:1.We analyze the reasons why the traditional attention mechanism has limitations for obtaining object location information,then we improve the traditional attention model used in convolutional neural networks.The method redesigns the calculation method of the attention probability map to reduce the numerical interval between the maximum probability and the minimum probability,which makes the probability map more smooth and more suitable for obtaining the pixel-by-pixel position information of the target object.In single target localization experiments on ImageNet dataset,this method is superior to related localization methods appeared in recent years,which indicating that the attention model can effectively obtain good target location information.2.Based on the characteristics of this attention model,a background area positioning method is proposed.Besides,post-processing is used to obtain the object position information to generate pixel-by-pixel semantic annotations,which are used to train image semantic segmentation model.In the experiment of large-scale semantic segmentation dataset PASCAL VOC2012,the result is superior to the majority of weakly supervised semantic segmentation methods and the accuracy increased by 1.3% ? 3.6% using only image-level category annotations,proving the effectiveness of the method.