摘要
现阶段图像美学评估存在两个问题,首先网络模型复杂度高、参数量大;其次分别考虑图像的局部(细粒度)特征和全局特征,忽视其美学相关性。为使高性能的美学评价模型在移动端部署成为可能,通过对比不同轻量化主干网络对实验效果的影响,选取性能最优的轻量网络MobilenetV2作为基准网络,融合图像的浅层(局部)和深层(全局)特征进行特征增强,提高网络模型对美学特征的理解。与先进美学评价模型相比,性能提升且参数量大幅减少,为美学评价模型在边缘设备上的部署提供了新的研究思路。
There are two problems in image aesthetics evaluation at this stage.Firstly the network model has high complexity and a large number of parameters.Secondly the local(fine-grained)and global features of an image are considered separately,and their aesthetic relevance is ignored.In order to make it possible to deploy a high-performance aesthetic evaluation model on the mobile terminal,by comparing the influence of different lightweight backbone networks on the experimental results,the lightweight network MobilenetV2 with the best performance is selected as the benchmark network,and the shallow(local)and Deep(global)features are used for feature enhancement to improve the network model's understanding of aesthetic features.Compared with the advanced aesthetic evaluation model,the performance is improved and the number of parameters is greatly reduced,which provides a new research idea for the deployment of the aesthetic evaluation model on edge devices.
作者
李健
王文文
李恩
LI Jian;WANG Wenwen;LI En(School of Electronic Information and Artificial Intelligence,Xi'an Shanxi 710021;School of Design and Art,Shaanxi University of Science and Technology,Xi'an Shanxi 710021)
出处
《软件》
2023年第4期94-101,共8页
Software
关键词
图像质量评估
计算机图像处理
计算机美学评估
计算机美学质量评价
image quality assessment
computer image processing
computer aesthetics assessment
computer aesthetics quality assessment
作者简介
李健(1975-),男,陕西渭南人,博士,教授,研究方向:计算机视觉、数字孪生;王文文(1997-),女,陕西咸阳人,硕士在读,研究方向:计算机图像美学评价;李恩(1997-),男,陕西渭南人,硕士在读,研究方向:艺术设计。
