图像质量评价（6）美学评估总结

忘是亡心i 2022-02-24 08:34 686阅读 0赞

参考博客：

[https://blog.csdn.net/God\_68/article/details/81534845][https_blog.csdn.net_God_68_article_details_81534845]

参考微信公众号：

[https://mp.weixin.qq.com/s?\_\_biz=MzI4MjU4MzAyMg==&mid=2247483714&idx=1&sn=6ea19697fb803069b1726b19e65e60b3&chksm=eb968bc3dce102d5353af9b71cae1433b3e1d4b77d86369f58096de8239e9b1c307949fb7960&mpshare=1&scene=23&srcid=0730nEt1dCd4kt3B5vkmtt2E\#rd][https_mp.weixin.qq.com_s_biz_MzI4MjU4MzAyMg_mid_2247483714_idx_1_sn_6ea19697fb803069b1726b19e65e60b3_chksm_eb968bc3dce102d5353af9b71cae1433b3e1d4b77d86369f58096de8239e9b1c307949fb7960_mpshare_1_scene_23_srcid_0730nEt1dCd4kt3B5vkmtt2E_rd]

一、数据库和目前现状

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MTc3MDE2OQ_size_16_color_FFFFFF_t_70][]

目前图像美学评估：最好的数据集就是AVA，有25万+

其他的即使加上也没什么效果，因此美学的如果要优化，就要标注数据集。

如果要特别好的话，需要标注类似于AVA的10个等级的数据，很难。

美学数据集：

AVA（2012）：255530幅，1-10分。每幅图被评分的人数范围为78-539，平均210。

200多篇论文引用

AADB（2016）：8种美学因素（平衡、颜色和谐型、趣味、景深、光照、主体、三分、颜色丰富性）的二值分类评价（在单个美学因素上的“好”与“不好”）。因素评价太多简单。

PCCD（2017）：评分+分布+多人语言评论，但太少4307幅

AROD（2018）：38万幅，图像查看和点赞次数计算得到美学评分

二、美学评估算法

1、NIMA：google，2017

训练集：AVA

思路：将质量评价当前计算1-10类的分类分布

论文：

[https://arxiv.org/abs/1709.05424][https_arxiv.org_abs_1709.05424]

github资源：

（1）1个模型

非官方复现代码：

[https://github.com/titu1994/neural-image-assessment][https_github.com_titu1994_neural-image-assessment]

最推荐：NasNet效果

[https://github.com/tfriedel/neural-image-assessment][https_github.com_tfriedel_neural-image-assessment]

nasnet，mobilenet

（2）2个模型：美学，技术

[https://github.com/idealo/image-quality-assessment/tree/master/models/MobileNet][https_github.com_idealo_image-quality-assessment_tree_master_models_MobileNet]

mobilenet

相关：

[https://blog.csdn.net/qq\_22194315/article/details/82713283][https_blog.csdn.net_qq_22194315_article_details_82713283]

NIMA解释

[http://server.chinabyte.com/125/14395125.shtml?S=9rrf3v9huog][http_server.chinabyte.com_125_14395125.shtml_S_9rrf3v9huog]

[https://cloud.tencent.com/developer/news/49602][https_cloud.tencent.com_developer_news_49602]

[https://blog.csdn.net/yH0VLDe8VG8ep9VGe/article/details/78849545][https_blog.csdn.net_yH0VLDe8VG8ep9VGe_article_details_78849545]

[https://blog.csdn.net/qq\_22194315/article/details/82713283][https_blog.csdn.net_qq_22194315_article_details_82713283]

知乎链接：

[https://zhuanlan.zhihu.com/p/33194024][https_zhuanlan.zhihu.com_p_33194024]

2、ILGNet：北京电子科技，2016

论文：

[http://jinxin.me/downloads/papers/019-WCSP2016a/ILGNet-Final.pdf][http_jinxin.me_downloads_papers_019-WCSP2016a_ILGNet-Final.pdf]

github：

[https://github.com/BestiVictory/ILGnet][https_github.com_BestiVictory_ILGnet]

caffe，模型大小：500M

主要思路：将质量评价当作分类，GoogleNet模型做fine-tune

增强数据集：AVA

三、我的思考：

美学我打算直接用，不打算做个性化了。

标数据是个很大的问题。

[https_blog.csdn.net_God_68_article_details_81534845]: https://blog.csdn.net/God_68/article/details/81534845
[https_mp.weixin.qq.com_s_biz_MzI4MjU4MzAyMg_mid_2247483714_idx_1_sn_6ea19697fb803069b1726b19e65e60b3_chksm_eb968bc3dce102d5353af9b71cae1433b3e1d4b77d86369f58096de8239e9b1c307949fb7960_mpshare_1_scene_23_srcid_0730nEt1dCd4kt3B5vkmtt2E_rd]: https://mp.weixin.qq.com/s?__biz=MzI4MjU4MzAyMg==&mid=2247483714&idx=1&sn=6ea19697fb803069b1726b19e65e60b3&chksm=eb968bc3dce102d5353af9b71cae1433b3e1d4b77d86369f58096de8239e9b1c307949fb7960&mpshare=1&scene=23&srcid=0730nEt1dCd4kt3B5vkmtt2E#rd
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MTc3MDE2OQ_size_16_color_FFFFFF_t_70]: /images/20220224/463c433faf874c24a6e360d678391c87.png
[https_arxiv.org_abs_1709.05424]: https://arxiv.org/abs/1709.05424
[https_github.com_titu1994_neural-image-assessment]: https://github.com/titu1994/neural-image-assessment
[https_github.com_tfriedel_neural-image-assessment]: https://github.com/tfriedel/neural-image-assessment
[https_github.com_idealo_image-quality-assessment_tree_master_models_MobileNet]: https://github.com/idealo/image-quality-assessment/tree/master/models/MobileNet
[https_blog.csdn.net_qq_22194315_article_details_82713283]: https://blog.csdn.net/qq_22194315/article/details/82713283
[http_server.chinabyte.com_125_14395125.shtml_S_9rrf3v9huog]: http://server.chinabyte.com/125/14395125.shtml?S=9rrf3v9huog
[https_cloud.tencent.com_developer_news_49602]: https://cloud.tencent.com/developer/news/49602
[https_blog.csdn.net_yH0VLDe8VG8ep9VGe_article_details_78849545]: https://blog.csdn.net/yH0VLDe8VG8ep9VGe/article/details/78849545
[https_zhuanlan.zhihu.com_p_33194024]: https://zhuanlan.zhihu.com/p/33194024
[http_jinxin.me_downloads_papers_019-WCSP2016a_ILGNet-Final.pdf]: http://jinxin.me/downloads/papers/019-WCSP2016a/ILGNet-Final.pdf
[https_github.com_BestiVictory_ILGnet]: https://github.com/BestiVictory/ILGnet