Classification oriented embedded image coding
Shaorong Chang · 2004
This paper discusses the efficient compression of images to improve the classification associated with the decoded imagery. The set partitioning in hierarchical trees (SPIHT) algorithm, an efficient wavelet-based progressive image-compression scheme, was originally designed to minimize the mean-squared error (MSE) between the original and decoded imagery. The image is first segmented at the encoder by an unsupervised method using a hidden Markov tree (HMT) mixture model in the wavelet domain. By using the kernel matching pursuits (KMP) method the recognition importance of each wavelet subband is estimated. By comparison using synthesized data, the compression and classification performance of the modified SPIHT algorithm is comparable to Bayes TSVQ, along with the advantages of fast speed and no requirement of codebook design and possibly transmission.