Vision-based image similarity measurement for image search similarity

In various applications across different platforms, image similarity features such as image searching and similar image recommendations are widely used. However, the challenges of semantic gap and querying speed continue to pose significant challenges in image similarity searching. In this study, we propose a novel solution to address these issues using contrastive learning within the TensorFlow Similarity library. Specifically, we trained and tested our proposed method using the Caltech-256 dataset and further evaluated it on the Corel1K dataset. Our work distinguishes itself from previous studies that primarily focus on evaluating accuracy while neglecting the importance of speed evaluation. As such, we propose evaluating both the mean average precision score and query time spending. Our experimental results reveal that our method based on EfficientNet (B7) yields the best average precision scores of 0.93 on the Caltech-256 test dataset and 0.94 on the Corel1K dataset. However, other methods achieve faster query times, although their average precision scores are significantly lower. © 2023, The Author(s), under exclusive licence to Bharati Vidyapeeth's Institute of Computer Applications and Management.