Trademark Image Recognition

Introduction

The aim of the PROFI project is to invent and develop new techniques for the retrieval of figurative images (such as clip art, logos, signs) from large databases. The techniques are based on the extraction and matching of perceptually relevant shape features, thereby overcoming many of the limitations of existing figurative image retrieval methods. The project is developing new algorithms and systems for:

• Perceptual segmentation of raw images, and grouping of shape elements.
• Matching of geometrical patterns representing shape features.
• Indexing shape features in huge databases of figurative images.
• Indexing the relative spatial layout of shape features within these images.

We have developed a trademark retrieval system as part of the EU FP6 project - PROFI.  The York work has focussed on analysing how humans break down figurative images and developing tools to mimic the human break down. One tool allows us to generate multiple representations of trademark images to allow the images to be processed at different perceptual levels.  A second tool comprises: edge detection followed by edge segmentation and then shape finding.  The shape finding uses the edge segments and a graph searching technique to construct the shapes. Finally, we aim to index and match trademark images using a fast graph matching framework which uses the edge segments and shapes to match the trademark images at different perceptual levels.

Problem

Figurative image databases such as trademark databases frequently contain millions of images.  The PROFI project aims to index such image databases by using the elements (shapes) identified within each image as the keys to the index.  The PROFI perceptual segmentation algorithm generates the shapes for indexing the images in the database. 
The perceptual segmentation of raw images is computationally intensive.  Each image may take 5 minutes or more to segment.  PROFI image segmentation thus requires a highly parallel machine to allow the entire image database to be processed in a timely manner. 

System

The system chosen for the image segmentation was the York node of the White Rose Grid which comprises a cluster of dual and quad processor machines with a total of 60 AMD Opteron 275 dual core processors (120 CPU cores total) and total memory across all nodes of 288GB. This is supported by a 6TB GPFS file system. This system was chosen as it provides a highly-parallel processing environment allowing many images to be processed in parallel.

Results

The WRG was able to process 11,500 raw images using perceptual segmentation in 1 day.  Using a single-processor desktop machine would take a minimum of 40 days (assuming no downtime) to process the same database of images.