Pattern recognition by grouping areas in DCT compressed images
Daidi Zhong, Irek Defee · Nordic Signal Processing Symposium · 2004
Images and video are almost exclusively handled in compressed formats based on quantized block DCT transform. Information extraction from images and video has been traditionally studied in the pixel domain. At present methods operating in the DCT domain are more natural and required. There is also argument for DCT based information extraction based on efficiency: compressed images preserve perceptually relevant information at greatly reduced size. This means that all perceptually non-relevant information is eliminated which should facilitate information extraction. While there have been some investigations of pattern recognition in compressed domain in the past, in this paper we analyze the problem from the compression and information reduction perspective. Pattern recognition method based on optimized quantization of DCT blocks and density of blocks in regions is introduced and illustrated on the example of face detection and recognition problem.