Detection of Near-Duplicate Images using Statistical Texture Features

With advancements in digital image processing technology, it is tremendously effortless for anybody to reacquisition high-quality multimedia from an imitating medium. Subsequently, digital image forensics has become progressively significant, in which photographic copying detection is one emerging field. The re-acquisition of digital media results in the introduction of artifacts. The visual quality dropping or blurriness is one of them, which can be effectively extracted using the texture features. State-of-the-art near-duplicate forensic methods are based on either some complex dictionary learning procedure or an improvisational group of features. In contrast to that, we proposed a model based on texture-based statistical features, namely, Multi-scale Local Binary Pattern (MLBP), the Log-Gabor Filter (LoG), and Dual-Tree Complex Wavelet Transform (DTCWT) for classification of near-duplicate or re-acquire images from the singly captured images. The adopted features are second-order statistical texture features. The extracted features perform well on the publicly available datasets, NTU-ROSE, ICL, and Mturk datasets. The comparisons with existing methods exhibit the computational accuracy and efficiency of the proposed model. Features extracted from the dataset of length 32 allows an SVM classifier to achieves an accuracy of 99% for the NTU-ROSE, 96% for ICL and 90% for the Mturk dataset.