A Multitask Convolutional Neural Network for Artwork Appreciation

Abstract

The computational aesthetics of pictorial art is an important part of human artistic creation, and the computational aesthetics of pictorial art images is a computationally computable human aesthetic process using machines, which has important applications and scientific significance in the automated analysis of large-scale paintings and the computational modeling of perception by machines. To this end, this paper proposes a multitask convolutional neural network model for emotion and rating of artworks. (1) An artwork appreciation dataset consisting of fifty Chinese paintings and fifty Western oil paintings was created, and twenty subjects were recruited to score the art appreciation of one hundred artworks in the dataset, covering both painting aesthetic evaluation and painting emotion evaluation. (2) Based on the artwork art appreciation dataset, an AlexNet-based convolutional neural network model is proposed to utilize the powerful feature extraction and classification capabilities of neural networks to complete artwork art appreciation, and an oversampling method and multitask learning method are used to improve the overall recognition accuracy. (3) Compared with the combination of traditional manual features + machine learning algorithms, the end-to-end multitask convolutional neural network proposed in this paper has the highest accuracy rate of 74.57%/71.43%/74.12%.