摘要
针对小样本语义分割中同类别支持图像与查询图像存在外观差异较大的问题,提出融合高斯过程的自支持匹配小样本语义分割模型。提出的模型在自支持匹配小样本语义分割模型的基础上,首先融入高斯过程,对分布在深层特征空间上的复杂外观进行建模,捕获更多空间细节信息来表示数据分布;随后设计特征增强模块,在空间层对支持特征与查询特征进行信息交互,在通道层进行注意力加权,进一步增强相同类之间的全局相似性,捕获更多目标类别信息;最后利用Gram矩阵量化支持图像和查询图像之间外观差异的大小,从而融合原型匹配的结果,产生更准确的分割图像。实验结果表明:与现有方法相比,所提模型在更强的主干网络下具有较好的分割结果和更少的参数量,在5-shot的设定下,所提模型在PASCAL−5i数据集上平均交并比(mean Intersection over Union,mIoU)达到最优值,提升了0.4%;在COCO−20i数据集上的子集mIoU取得最优值,分别提升了2.2%和1.0%,表明该模型的有效性和先进性。
Aiming at the problem of large difference in appearance between the same category support image and the query image in few-shot semantic segmentation,a self-support matching few-shot semantic segmentation model based on Gaussian process is proposed.Based on the self-support matching few-shot semantic segmentation model,the proposed model first integrates the Gaussian process to model the complex appearance distributed on the deep feature space,and captures more spatial details to represent the data distribution.Then,a feature enhancement module is designed to carry out information exchange between supporting features and query features in the spatial layer,and carry out attention weighting in the channel layer to further enhance the global similarity between the same classes and capture more target category information.Finally,the Gram matrix is used to quantify the size of the difference in the appearance of the support image and query image,and fuse the results of prototype matching to produce more accurate segmented images.The experimental results show that compared with the existing methods,the proposed model has better segmentation results and fewer parameters under a stronger backbone network,and the mean Intersection over Union(mIoU)of the proposed model reaches the optimal value on the PASCAL−5i dataset under the setting of 5-shot,which is increased by 0.4%.The optimal values of subsets mIoU on the COCO−20i dataset are improved by 2.2%and 1.0%,respectively,indicating the effectiveness and advanced nature of the model.
作者
罗余特
宣士斌
张慧
刘成星
LUO Yute;XUAN Shibin;ZHANG Hui;LIU Chengxing(School of Artificial Inteligence,Guangxi Minzu University,Nanning 530006,China;Guangxi Key Laboratory of Hybrid Computation and IC Design and Analysis,Nanning 530006,China)
出处
《微电子学与计算机》
2024年第8期62-72,共11页
Microelectronics & Computer
基金
国家自然科学基金(61866003,62062011)。
关键词
小样本语义分割
原型结构
自支持匹配
高斯过程
信息交互
few-shot semantic segmentation
prototype structure
self-support matching
Gaussian process
information interaction
