Adaptive Weighted Matching of Deep Convolutional Features for Painting Retrieval

We focus on painting retrieval problem, and our motivation is to find out similar paintings and assist painting plagiarism identification. Similar painting retrieval is much more challenging than natural image retrieval, since different paintings have different styles and the similarity of paintings is difficult to measure. In this paper, we define the similarity of paintings from the perspectives of both semantics and structure rather than pixel level color, texture and shape. Specifically, we use the pooling activations of Convolutional Neural Network (CNN) to represent painting features, which preserves both semantic information and structure information. We propose an adaptive weighted matching approach to measure the similarity of paintings, and embed it into a painting retrieval framework. Furthermore, we propose a feature map selection approach to reduce redundancy based on the weights. We collect a new paintings dataset to evaluate painting similarity, which consists of 324 query-reference image pairs from China Artists Association, and 7200 distracted painting images from the Internet, which contains the most common similarity cases. Our approach obtains promising results on the dataset, confirming the superiority of our approach.