We propose a closed-form rate allocation scheme (RAS) for entropy-coded uniform scalar quantization of dependent sources in classification problems. The proposed RAS is applicable to nonbinary classification with piecewise monotonic unquantized Bayes decision boundaries. The RAS is also extended to joint compression and classification.

We present a study of image features for cancer diagnosis and Gleason grading of the histological images of prostate. In diagnosis, the tissue image is classified into the tumor and nontumor classes. In Gleason grading, which characterizes tumor aggressiveness, the image is classified as containing a low- or high-grade tumor. The image sets used in this paper consisted of 367 and 268 color images for the diagnosis and Gleason grading problems, respectively, and were captured from representative areas of hematoxylin and eosin-stained tissue retrieved from tissue microarray cores or whole sections. The primary contribution of this paper is to aggregate color, texture, and morphometric cues at the global and histological object levels for classification. Features representing different visual cues were combined in a supervised learning framework. We compared the performance of Gaussian, k-nearest neighbor, and support vector machine classifiers together with the sequential forward feature selection algorithm. On diagnosis, using a five-fold cross-validation estimate, an accuracy of 96.7% was obtained. On Gleason grading, the achieved accuracy of classification into low- and high-grade classes was 81.0%.

The widespread adoption of the JPEG2000 standard calls for the development of computationally efficient algorithms to analyze the content of imagery compressed using this standard. For this purpose, we propose the use of the information content (IC) of wavelet subbands, defined as the number of bytes that JPEG2000 spends to encode the subbands. The IC of subbands can be obtained from the packet headers of the JPEG2000 codestream, thereby avoiding decompressing the arithmetically encoded bitplane data. We present experimental results for two content analysis tasks; namely, image classification and scene change detection. Our results indicate that performance comparable to that of methods operating on decompressed data can be achieved, while saving computational and bandwidth resources.