soruly

New Image Search Algorithm (JCD)

Added 2020-04-13 10:48:27 +0000 UTC

A new search algorithm has been added to trace.moe. You can try this using the "Use new algo" check box near the search button. This new algorithm has better support on flipped and cropped images. See results above.

This new algorithm still needs fine-tuning of its parameters, so I'm now opening this option for everyone to try. For developers using the API, you can use /search?method=jc . In upcoming months, I'll review the performance and accuracy of these two algorithms and see if JCD is good enough to replace ColorLayout. Or else, methods to combine the two.

The new algorithm, JCD (Joint Composite Descriptor) is composed of CEDD (Color and Edge Directivity Descriptor) and FCTH (Fuzzy Color and Texture Histogram), takes both color, edge (shape) and texture into image analysis. trace.moe has been using ColorLayout descriptor, which does not analyze the image edge and shape but only the distribution of colors in a 8x8 grid. With this new algorithm, it makes it possible to search for flipped/cropped images which previously fails.

Note that none of these Image Descriptors are invented by me. If you're interested in the principles of the the algorithm(s), please read the paper from the original author. and its open source implementation, LIRE.

When indexed on the same video data set, the size of indexed database (extracted) is 232GB, 85% larger than that of ColorLayout (125GB). The database cache of two algos and memory for apache solr now occupies 420GB RAM (out of 512GB) RAM on server.

Serving the database for the new algorithm requires powerful servers. Please continue to support this project for ACG fans! :3