Duplicate Image Representation Based on Semi-Supervised Learning

Introduction

With the explosive growth of multimedia data on the Internet (Xiao, Cheng, Wei, Li, Wang, & Xu, 2019; Xiao, Cheng, & Liu, 2019), duplicate images are becoming more and more common, which brings new challenges to many image applications (Yan, Sethi, Rangarajan, Raju, & Ranka2017; Wang, Jiang, & Su, 2018; Chen, Li, Zhang, Hsu, & Wang, 2017). For example, in an image retrieval system, by high performance computing (Amsaleg, Gudmundsson, Jonsson, & Franklin, 2018), people return various related images according to a query image. If the returned images contain a large amount of duplicate content, it will bring great inconvenience to user browse. In additional, due to the efficiency and convenience of multimedia data acquisition, images can be easily tampered and copied, which brings many problems for copyright protection. Therefore, there is an urgent need for duplicate image detection technology, which can quickly, accurately, reliably find all copies of a given image on the Internet.

The traditional digital watermarking technology can effectively detect duplicate images by embedding additional information in the images (Gao, Chi, Du, & Diao, 2017; Prasanth, & Chandra, 2018). However, this method must embed watermark information into images before images are released, which is powerless for the published images. This characteristic greatly limits the application of digital watermark. Different from the traditional digital watermark technology, the content-based duplicate image detection technology can extract the invariant features of images to represent images and utilize the invariant features to detect duplicate images (Hyunwoo, SungRyull, & Junmo, 2019; Wang, & Zhou, 2018).This method not only has a good discrimination effect, but also maintains image content. So it has better applicability. According to the selected areas, the invariant features can be divided into two categories: global features and local features. Although global features have better retrieval efficiency and scalability, they are powerless for geometric transformations, such as cropping, rotation, and translation. Therefore, in order to solve this problem, more robust local features are needed. Compared with global features, local features are more robust to optical and geometric transformations. So they are widely used in duplicate image detection. However, local features represent an image as a set of high-dimensional feature vectors. This brings the great pressure on the organization and management of the data.

In recent years, the Bag-of-Feature (BoF) model is considered to be a key technology to large-scale duplicate image retrieval (Sivic, & Zisserman, 2003; Zhou, Narentuya, Tang, & Liu, 2018; Ulutas, Ustubioglu, Ulutas, & Nabiyev, 2018).The BoF model represents each image as a frequency histogram of visual words, which are obtained by quantify local descriptors to the closest cluster centroids as visual words. Then fast access to the frequency histograms is implemented by an inverted indexing. Because the quantization technique avoids storing and comparing high-dimensional vectors, it can effectively reduce memory usage and improve retrieval efficiency. However, quantization process greatly reduces the discriminability of image representation. This is because the construction of visual dictionary in quantization process is mainly implemented by K-means algorithm. When the dictionary size exceeds 10⁵, the training is difficult. The size limit of visual dictionary reduces the discriminative powers of visual words. Although researchers have proposed some optimization algorithms for this problem, such as hierarchical K-means (Ulutas, Ustubioglu, Ulutas, & Nabiyev, 2018) and approximate K-means (Arai, & Barakbah, 2007), the memory limitation is still a fundamental reason that restricts the clustering effect.