A Fine Image Segmentation Algorithm Driven By Detailed Semantics

In recent years,due to the constant advances in deep learning of network performance,image segmentation algorithm achieved a breakthrough.While existing algorithm can better segment the target object,but there is still a certain lack of detail on the edge.Image segmentation is good or bad,often affect the subsequent image processing tasks.Therefore,developed an accurate image segmentation algorithm and application of major significance in promoting the overall visual computing technologies.Aimed at the poor effects of existing image segmentation algorithms in detecting the slender and sharp parts of an object,this paper proposes a fine image segmentation framework driven by detailed semantics.The detailed semantic extraction model extracts the detailed semantics of labels and increases network branches aimed at semantics details,which enables the network to be more focused on the detailed parts of objects without losing the semantics of key parts of objects,so that a finer segmentation can be realized.The specific research results are as follows:(1)The model adds a channel attention mechanism to Mask R-CNN Feature Pyramid and enhances the dependency between different channels,so that the network is better able to extract features.(2)This paper introduces the concept of detailed semantics and input the detailed information of objects as labels into the network,which enhances the network segmentation accuracy in key semantic regions.This paper also discusses several forms of detailed semantics,including cornor semantics,profile semantics,axis region semantics.Traditional methods are used to realize the extraction of different semantics.(3)This paper builds an end-to-end image instance segmentation algorithm framework.By using a dual-task supervision method,the original segmentation task and the detailed semantic supervision task are mutually guided and promoted.This paper also tests the superiority of this algorithm in standard COCO dataset and self-made indoor-scene datasets.The proposed fine image segmentation framework driven by detailed semantics introduces attention mechanism and detailed semantics,so it is better able to extract features,learn detailed semantics and thus realize a finer segmentation than existing instance segmentation techniques.In addition,the model proposed in this paper has a larger applied range,which can not only be applied to Mask R-CNN but also used in any other image segmentation algorithms.