Research On The Art Image Classification Algorithm Based On Deep Learning

With the advances of human living standard and the improvement of art appreciation,more and more people have invested in the creation and collection of art images.At the same time,the development of next-generation information technologies such as big data and artificial intelligence has also boosted art creation methods more diverse and art styles richer.However,in the face of an increasing and huge number of art images,the current classification management work is mainly implemented manually by rare professionals,and the cost in term of manpower and financial resources is very large.Hence,it is of great significance to research on how to classify various types of art images to help users filter images that better meet their needs efficiently and accurately.To match the requirements of efficient and fast classification or label of artistic images,this thesis sets research objects as printmaking,national painting,oil painting,water painting,and watercolor painting,and proposes an algorithm for the classification of art images based on double kernels squeeze and excitation neural network,and then implements them systematically.The main work of this thesis consists of the following three-fold:(1)Proposed an art image classification algorithm.In order to extract the overall style features and local detail features of above mentioned five types of art images fully,this thesis proposes a double kernels squeeze and excitation module(i.e.,DKSE).The convolutions of different sizes in the module realize the extraction of overall image features and local detail features,squeeze and excitation operations achieve the enhancement of main features and the suppression of irrelevant features,and use the DKSE modules and depthwise separable convolution modules to build a deep convolutional neural network to achieve classification of five types of art images.The proposed algorithm in this thesis can effectively extract the overall features and local details of artistic images,and the accuracy rate of artistic images classification is 87.58%.(2)The validation and analysis of proposed algorithm.Firstly,the effects of image data enhancement operations on the classification results were verified by many experiments,and then the effects of the parameters in the double kernels squeeze and excitation module on the model classification results were obtained,and a set of reasonable configuration parameters was obtained.Afterwards,Grad-CAM(Gradient-weighted Class Activation Mapping)algorithm visualizes the image feature areas of interest in the network model in the form of a heat map to illustrate the effectiveness of the algorithm feature selection in this paper.The double kernels squeeze and excitation module is improved,and the improved module reduces the network model training time,and accuracy is improved.Finally,this network model and other network models are used to classify 3,4,and 6 art images to verify the effect of image category number on classification performance of network model,respectively.The experimental results show that compared with other mainstream network models and traditional classification algorithms,the accuracy rate of art image classification in this algorithm is higher than the current mainstream CNN models.Compared with traditional art image feature extraction methods,it does not rely on researchers to extract image features.The extracted features are more comprehensive and more accurate.(3)Design and implementation of intelligent classification system for art images.The system uses B/S architecture and is built by using Python and Django framework.In actual use,the user uploads the art image to be classified to the server,and the server returns the result classified based on the algorithm in this article to the front end in real time.The time taken by the back-end network model algorithm from receiving the image to returning the classification result is about 1.4s.