Research On Ink-wash Painting Classification Method Based Upon Deep Feature Fusion

Chinese ink-wash painting is a unique style of painting in China,which has been passed down for thousands of years,and it is also a treasure in the palace of human art.With the increasing demand for automatic appreciation and identification of art works,how to accurately extract features and classify the digital images of ink-wash paintings has become a hot research issue at present.In the early research work,manual design features were used to depict the characteristics of ink-wash paintings,but this manual design feature extraction method has great limitations.With the rapid development of deep learning technology,some methods using deep neural network frameworks to extract advanced features of ink-wash painting images have emerged,and the performance of classification task has been improved accordingly.On the basis of previous work,this thesis carries out in-depth research on feature fusion and structure of deep neural network model.Based on the excellent learning ability of deep learning in image feature extraction,this thesis combines optimization strategies to conduct research on feature extraction and fusion of ink-wash paintings,so as to improve the classification performance.This thesis mainly includes the following four aspects:(1)The classification research of ink-wash paintings based on discrete cosine transform and convolutional neural network was carried out,which significantly improved the classification performance of the two painting artistic styles of Gongbi and Xieyi.In order to solve the problem of low classification performance caused by manual design feature representation in the early classification of ink-wash paintings,a model based on discrete cosine transform and convolutional neural network was proposed.The model takes full advantage of the characteristics of discrete cosine transform and convolutional neural network,inputs the extracted feature into the designed neural network for fusion,and combines support vector machine for classification.Finally,sufficient experiments were carried out on the data set of Gongbi and Xieyi classification constructed in this thesis to verify the validity of the model.(2)The deep neural network model research based on heterogeneous feature fusion was carried out to reduce the influence of the diversity of ink-wash painting artistic feature styles on classification.Due to the diversity of artists’ painting styles,the ink-wash paintings have complex mixed characteristics of isomorphism and heterogeneity.Aiming at this problem,this thesis proposed a neural network model based on heterogeneous feature fusion,which effectively improves the feature fusion effect in the mixed problem of multi-feature isomorphism and heterogeneity,and strengthens the heterogeneous features that are more helpful for classification.Finally,the experimental results on the data set of ink-wash paintings for the multi-classification of writers constructed in this thesis also verify that the model can effectively improve the classification performance of the works of multiple writers with similar artistic styles.(3)The model research based on the multiple high-level semantic features fusion was carried out to improve classification performance from the features spatial level of semantic and artistic style.Through the research,we found that it is difficult to extract low-level features from the image that are very effective for classification for this special art form of ink-wash painting.However,there are often better results in the high-level semantic feature space,so a classification model of ink-wash paintings based on the fusion of two-channel semantic features was proposed.By designing two feature learning branches,the model extracts the high-level semantic feature information of the image and the high-level artistic feature information of the brush strokes based on the Dense Net network respectively.Finally,it was verified on the data set we constructed,and the performance exceeded the existing methods.(4)The research on multi-level attention and multi-scale feature fusion was carried out to improve the effect of feature selection and fusion at different levels and classification performance.Aiming at the problem that feature weights are not optimized in most existing works,a network model based on multi-level attention and multi-scale feature fusion was proposed.By extracting low-level,middle-level and high-level features,and using spatial and channel attention mechanisms and feature fusion strategies,the model effectively improves the feature weights that contribute most to the classification task,reduces the feature weights with high feature space overlap rate,and performs the classification task.By comparing the ink-wash painting data set we constructed with other methods,it is verified that the method in this thesis is superior to other recent works.The above four research works are aimed at how to effectively extract and integrate the most helpful features for the current classification task under the circumstance that the artistic features of ink-wash painting are varied and difficult to distinguish.They are compared with other methods on the ink-wash painting data sets we constructed,which verifies the effectiveness of the work of this thesis.