The Novel Instance Segmentation Method Based on Multi-Level Features and Joint Attention  

作　　者：XU Bowen LU Yinan WU Tieru GUO Xiaoxin 

机构地区：[1]College of Computer Science and Technology,Jilin University,Changchun 130012,China [2]College of Mathematics,Jilin University,Changchun 130012,China

出　　处：《Chinese Journal of Electronics》2023年第5期1160-1168,共9页电子学报（英文版）

基　　金：supported by the National Natural Science Foundation of China(61872162,82071995);the Key Research and Development Program of Jilin Province(20210301001GX,20220201141GX);the Natural Science Foundation of Jilin Province(20200201292JC).

摘　　要：Instance segmentation is an important task in computer vision.In order to enhance the multi-level features expression ability of the segmentation networks,a novel module is proposed in this paper.Firstly,we design a weighted bi-directional feature fusion way by computing the weight distribution function of bi-directional feature pyramid network.Secondly,we propose a joint attention mechanism to effectively filter different levels of feature information by adopting serial and parallel ways to combine the channel attention and spatial attention modules.At the same time,the module uses dynamic convolution to stabilize the calculation speed while improve the 6.7%mean average precision of segmentation.The experiments on the COCO dataset demonstrate that the module can effectively improve the performance of the existing instance segmentation networks.

关 键 词：Instance segmentation Feature fusion Attention mechanism Dynamic convolution Deep neural network 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]