摘要
针对传统方法在古代壁画图像分割过程中出现的目标边界模糊、图像分割效率低等问题,提出一种基于PSPNet网络的多分类壁画图像分割模型(PSP-M).模型首先融合轻量级神经网络MobileNetV2,降低硬件条件对于模型训练的限制.其次通过全局金字塔模块,将不同级别的特征图拼接起来,避免了表征不同子区域之间关系的语境信息的丢失.最后利用金字塔场景解析网络嵌入壁画背景特征,减少特征损失的同时提高特征提取效率.实验结果表明,PSP-M模型较传统的图像分割模型在训练精确度上平均提升2%,峰值信噪比(PSNR)较实验对比模型平均提高1~2 dB,结构相似指标(SSIM)指标较实验对比模型平均提高0.1~0.2,实验验证了PSP-M模型在壁画分割方面的可行性.
In order to solve the problems of fuzzy boundary and low efficiency of traditional methods in ancient mural image segmentation,a multi classification mural image segmentation model based on PSPNet(PSP-M)is proposed.Firstly,the model integrates the lightweight neural network mobileNetV2 to reduce the limitation of hardware conditions for model training.Secondly,through the global pyramid module,the feature maps of different levels are spliced to avoid the loss of context information representing the relationship between different sub regions.Finally,the pyramid scene analysis network is used to embed the mural background features to reduce the loss of features and improve the efficiency of feature extraction.The experimental results show that the training accuracy of PSP-M model is 2%higher than that of the traditional image segmentation model,the peak signal-to-noise ratio(PSNR)of the model is 1-2 dB higher than that of the experimental comparison model,and the structural similarity index(SSIM)of the model is 0.1-0.2 higher than that of the experimental comparison model.The experimental results verify the feasibility of PSP-M model in mural segmentation.
作者
曹建芳
田晓东
贾一鸣
闫敏敏
马尚
Cao Jianfang;Tian Xiaodong;Jia Yiming;Yan Minmin;Ma Shang(Editorial College of Computer Science and Technology,Taiyuan University of Science and Technology,Taiyuan 030024,China;Computer Department,Xinzhou Teachers University,Xinzhou 034000,China)
出处
《河南师范大学学报（自然科学版）》
CAS
北大核心
2022年第4期65-75,共11页
Journal of Henan Normal University(Natural Science Edition)
基金
教育部人文社会科学研究规划基金(21YJAZH002)
山西省高等学校人文社会科学重点研究基地项目(20190130)。
关键词
壁画图像分割
金字塔池模块
深度可分离卷积
轻量级神经网络
mural image segmentation
Pyramid pooling module
depth separable convolution
lightweight neural network
