| Computational aesthetics is a discipline that studies how to use computer technology to judge and evaluate the aesthetics of images,videos,audios and other works of art,in order to evaluate the aesthetic value of works of art more objectively,accurately and efficiently.Chinese landscape painting is a cultural symbol of the Chinese nation,and the research and application of related computational aesthetics needs to be further developed.This paper focuses on the use of semantic segmentation technology in the field of computer vision to realize the visual evaluation of landscape painting,which has strong exploratory value.At present,in this field,the available open source data sets are extremely scarce,and there is a serious lack of systematic aesthetic evaluation;for semantic segmentation tasks,most of the existing image segmentation models have problems such as insufficient semantic extraction and insufficient use of features from different channels.Based on the design and creation of the Chinese landscape painting dataset,this paper carried out relevant innovative research and development work:1.Aiming at the problem that the shallow semantics and deep semantics of the model cannot interact,a cross-stage feature interaction model CSNet based on Swin Transformer is proposed,which makes the shallow effective semantics and deep semantics directly through the cross-stage feature filter and cross-stage feature interaction connection.interact.The experimental results show that the average accuracy of CSNet is significantly improved compared with other cutting-edge models,and the highest average accuracy reaches 91.47%.2.In view of the lack of correlation learning between pixels in the model and the slow reasoning speed,a non-local attention mechanism MCF Attention based on multi-channel fusion is proposed,which achieves low-level attention through multi-channel fusion modules and non-local attention mechanisms.Fine-grained association learning in pixels.The results of the ablation experiment show that MCF Attention improves the model inference speed from 1.7FPS to 2.5FPS while improving the accuracy to 91.78%.3.Aiming at the lack of aesthetic calculation systems for landscape paintings,a visual aesthetic evaluation system for Chinese landscape paintings was designed and implemented.The system uses React as the front-end development framework,Spring as the back-end development framework,MySQL as the database,combined with the CSNet model and the MCF Attention mechanism to realize the multi-dimensional visualization function based on image segmentation.Supports dimension evaluation and download of Chinese landscape paintings,and provides convenience for expanding Chinese landscape painting datasets through the image storage function.The results of function test,performance test,safety test and trial operation show that the system functions normally and runs well,and the expected results have been achieved. |
