A Multitask Convolutional Neural Network for Artwork Appreciation

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