Semantic Similar Image Search: A Command-Line Tool Based on CLIP Regular Research Paper, CSCI-RTPC

This paper introduces a practical and scalable approach for implementing a similar image search engine using Contrastive Language-Image Pre-Training (CLIP), OpenAl's model. The method provides a Command-Line Interface (CLI) and utilizes a cache layer powered by SQLite, a relational database, to enable efficient repetitive searches within large image databases. The effectiveness was evaluated on ImageNet-1k using a metric described in the paper, achieving 48.32 % accuracy. Scaling tests showed indexing time increases linearly with the number of images, while search time rose only slightly - for example, indexing over 1 million images took 26.36 times longer on an M1 Max CPU than indexing 50,000, but querying the larger set took just 2.75 times more. This approach would particularly benefit industries managing vast visual data volumes, such as media, security, manufacturing, and healthcare.