Performance Study of SIFT and Swin Transformer Algorithms in Fuzzy Duplicate Image Detection Problem

The quantity of data in any field of activity is increasing at an exponential rate on an annual basis. Images are no exception to this rule. In such circumstances, the importance of image analysis and processing is becoming increasingly evident. One such task is the detection of fuzzy duplicates of images, which is of significance for search engines and databases, as well as for fields such as medicine, security and law enforcement, advertising and marketing, and so on. Despite the existing variety of fuzzy duplicate search algorithms and models, there is no unambiguous answer as to which approach will be the most effective for a particular problem. This study examines two approaches to fuzzy duplicate image retrieval: one based on SIFT descriptors and the other on Swin Transformer. The current work presents the experimental results of the two models. The study compares the two models in terms of recall, specificity, PPV (positive predictive value), NPV (negative predictive value) and accuracy. The Swin Transformer model demonstrated superior performance across all evaluated metrics.