Keren Perlmutter

Image quality in lossy compressed digital mammograms

Abstract The substitution of digital representations for analog images provides access to methods for digital storage and transmission and enables the use of a variety of digital image processing techniques, including enhancement and computer assisted screening and diagnosis. Lossy compression can further improve the efficiency of transmission and storage and can facilitate subsequent image processing. Both digitization (or digital acquisition) and lossy compression alter an image from its traditional form, and hence it becomes important that any such alteration be shown to improve or at least not damage the utility of the image in a screening or diagnostic application. One approach to demonstrating in a quantifiable manner that a specific image mode is at least equal to another is by clinical experiment simulating ordinary practice and suitable statistical analysis. In this paper we describe a general protocol for performing such a verification and present preliminary results of a specific experiment designed to show that 12 bpp digital mammograms compressed in a lossy fashion to 0.015 bpp using an embedded wavelet coding scheme result in no significant differences from the analog or digital originals.

Text segmentation in mixed-mode images using classification trees and transform tree-structured vector quantization

Multimedia applications such as educational videos and color facsimile contain images that are rich in both textual and continuous tone data. Because these two types of data have different properties, segmentation of the images into text and continuous tone data can improve compression by allowing different compression parameters or even algorithms to be employed on the different types. We propose and compare algorithms that use classification trees (CLTR) or tree-structured vector quantization (TSVQ) for block-based classification in mixed-mode images. We also examine different types of features that can be used in these classifiers. The results show that using linear transform features with either the CLTR or TSVQ can be effective for accurate text classification. In addition, the results indicate that combining these classifiers with another TSVQ that is designed simultaneously to minimize both compression and classification error can provide better classification than does either system alone.

Training sequence size and vector quantizer performance

The authors examined vector quantizer performance as a function of training sequence size for tree-structured and full-search vector quantizers. The performance was measured by the mean-squared error between the input image and the quantizer output at a given bit rate. The training sequence size was measured either by the number of training images, or by the number of training vectors. When the training vectors were counted, they were selected randomly from among the training images. For every training sequence size, vector quantizers were developed from several different training sequences, and the distortion was calculated for different test sequences in a cross validation procedure. Preliminary results suggest that plots of distortion vs. number of training images follow an algebraic decay, as expected from analogous results of learning theory.<<ETX>>

Joint image classification and compression using hierarchical table-lookup vector quantization

Classification and compression play important roles today in communicating digital information and their combination is useful in many applications. The aim is to produce image classification without any further signal processing on the compressed image. This paper presents techniques for the design of block based joint classifier and quantizer classifiers/encoders implemented by table lookups. In the table lookup classifiers/encoders, input vectors to the encoders are used directly as addresses in code tables to choose the codewords with the appropriate classification information. In order to preserve manageable table sizes for large dimension VQs, hierarchical structures that quantize the vector successively in stages are used. Since both the classifier/encoder and decoder are implemented by table lookups, there are no arithmetic computations required in the final system implementation. They are unique in that both the classifier/encoder and the decoder are implemented with only table lookups and are amenable to efficient software and hardware solutions.

Evaluation of Bayes risk weighted vector quantization with posterior estimation in the detection of lesions in digitized mammograms

The automated detection of suspicious tissue in digital mammograms can provide a useful aid to diagnosis by permitting a radiologist to see all regions deemed suspicious by the computer. The authors apply to digital mammography a method that combines aspects of data compression techniques based on clustering and decision trees together with algorithms for classification and regression. The idea is to use a distortion measure in a clustering algorithm that includes both squared error for general appearance and average Bayes risk for classification accuracy. The algorithm structure is that of a vector quantization compression system that incorporates Bayes risk into the optimization algorithm.<<ETX>>

Tree-structured vector quantization with region-based classification

Unbalanced or pruned tree-structured vector quantization (PTSVQ), a variable-rate coding technique that tends to use more bits to code active regions of the image and fewer to code homogeneous ones, is developed based on a training sequence of typical images. A regression tree algorithm is used to segment the images of the training sequence using the x, y pixel location as a predictor for the intensity. This segmentation is used to partition the training data by region and generate separate codebooks for each region, and to allocate differing numbers of bits to the regions. Region-based classification requires no side information, as the decoder knows where in the image the current encoded block originated. These methods can enhance the perceptual quality of compressed images when compared with ordinary PTSVQ. Results for magnetic resonance data are shown.<<ETX>>

Spectral face clustering

Recognizing and clustering similar faces are important for organizing digital photos. We present a novel clustering method based on spectral clustering to group faces in a photo album. The main contribution is the proposal of a distance metric that is robust to outlier features present in the facial images.

Wavelet/TSVQ image coding with segmentation

The use of region-based coding is explored with a wavelet/TSVQ structure. Several methods of generating the segmentation map are discussed, including a recursive segmentation procedure that does not require any side information. The method is investigated on computerized tomographic chest scans, where the images are segmented into three regions-the background, the chest wall region, and the chest organs region. The background is considered of no importance, the chest wall region is considered of low importance, and the chest organs region is considered of high importance. At 0.20 bits per pixel, region-based coding provides a 2.0 dB improvement in the chest organs region at the expense of degradation in the clinically less relevant regions.

An iterative joint codebook and classifier improvement algorithm for finite-state vector quantization

A finite-state vector quantizer (FSVQ) is a multicodebook system in, which the current state (or codebook) is chosen as a function of the previously quantized vectors. The authors introduce a novel iterative algorithm for joint codebook and next state function design of full search finite-state vector quantizers. They consider the fixed-rate case, for which no optimal design strategy is known. A locally optimal set of codebooks is designed for the training data and then predecessors to the training vectors associated with each codebook are appropriately labelled and used in designing the classifier. The algorithm iterates between next state function and state codebook design until it arrives at a suitable solution. The proposed design consistently yields better performance than the traditional FSVQ design method (under identical state space and codebook constraints).<<ETX>>