Character Image Retrieval

A shape context based approach for fast matching Chinese characters in a document image. The experiment achieves 95% average precision and the average CPU time for matching two character images can be below 0.1 seconds without training on a 1.6 GHz processor.

Multilevel Thresholding & Document Layout Analysis

A cross-correlation approach combining Otsu's method for doing image segmentation can be applied to document image such as newspaper, magzine and other textual images. Image's foreground & background can be separated correctly, moreover, we prepare to separate image based on objects such as text regions, graphic regions and generic regions.

Document Image Compression

A new document image compression scheme is proposed based on Shape Context method. Compared to old standard such as JBIG and JBIG 2, our approach based on pattern matching can achieve more compression ratio and more functions such as: progressive transmission, content lossless compression and text region clustering.

EXIF Metadata for Photo Classification

We exploit camera EXIF metadata for picture scene classification. Possible scenes include Portrait, Outdoors, Sports, Indoors, Night, etc. The inference methods include Bayesian Networks and statistical KL divergence. The scene classification can be semantic and is superior to classification based on low-level features.