Research On The Author Recognition Of Fine-Art Paintings Based On Multi-Task Multi-Layer Feature Fusion Densenet

With the continuous development of digitalization,the intelligent analysis of paintings is becoming more and more important.The automatic author recognition of Fine-Art paintings will play an important role in the further development of the painting digitization.In recent years,deep learning represented by convolutional neural network has made good achievements in the field of image processing,so that we can efficiently solve the problem of automatic identification of the author of paintings.After a detailed analysis of the research in this field at home and abroad in recent years,this paper proposes a Multi-Task Multi-Layer Feature Fusion DenseNet and applies it to the task of identifying the author of paintings.In this paper,through transfer learning,improved DenseNet,Multi-Task Learning,and data augmentation,the model has a good feature learning ability and robustness in the case of a small number of datasets and many recognition categories.The main achievements of the paper are as follows:1.To solve the problem that the training images are insufficient for deep learning,this paper uses transfer learning.For the characteristics of painting recognition,lowlevel features such as texture,color,and shadow are of great significance for recognition.This paper proposes a densely connected convolutional network structure with multilayer feature fusion.Transfer learning uses the ImageNet dataset as a pre-training dataset,loads the pre-trained model parameters and then further fine-tune the model using the painting dataset.Multi-Layer Feature Fusion DenseNet(MFDN)is reconstructed based on the densely connected network.The feature maps are extracted before the transition layer of the original network.After separation,convolution,and pooling,they are directly connected in parallel to the final decision layer so that the deep and shallow information jointly make the final identify.The experiments show that the information learned from natural images can play a significant role in the identification of Fine-Art paintings.The recognition effect of MFDN is superior to the original densely connected network under multiple extraction ratios of shallow features.2.Aiming at the problem that the single-task learning ability to mine image features needs to be improved and the learned features are not robust,a multi-task learning method is proposed.In order to solve the problem that the convergence speed of related task is inconsistent with that of main task in multi task learning,a method of dynamically adjusting task weights in training process is proposed.The main task is to identify the author of the painting works,and the related task is to identify the style of painting works.The hard parameter sharing method is used to realize the information interaction between the tasks for multi-task-learning.Calculate the loss decline rate of the main task and the related task.When the loss decline rate of the main task is lower than the threshold than the related task,the weight of the related task is reduced and the weight of the main task is increased.Experiment results show that multi-task learning improves the learning ability of the model while reducing the possibility of learning redundant features.3.In response to the problem of how to continue to increase the model generalization ability in the case of limited data set,a method of using online data augmentation is proposed.The data enhancement factors used in online data enhancement include flip,rotation,zoom,projection,translation and the pictures that have been randomly processed by the enhancement factor are used as the real training set.Experiments prove that online data augmentation enhances the model's ability to cope with interference without increasing storage pressure,improves the robustness of the network,and has a better generalization ability.