Image segmentation and compression using hidden Markov models (Q2739558)

Image segmentation is a key step in image analysis and interpretation, specially in remote sensing. This is a process for dividing an image into its constituent parts, according to the users'goals. Image processing is becoming harder due to the huge volume of data currently available in the form of images (are getting larger and larger nowadays/ and image processing is getting harder). In fact, the huge amount of information is, in many cases, redundant. Therefore, image compressing has become a crucial field in image processing. and computationally fast and reliable techniques are needed to undertake the compressing process. NEWLINENEWLINENEWLINEThis book presents algorithms aimed at joint compression and segmentation. Both problems are also treated separately. The algorithms are based on context information using a hidden Markov model. In Chapter 2, background for statistical classification is provided. Optimal Bayesian classification procedures are analysed and algorithms based upon these techniques are reviewed. Markov processes are taken into consideration to model pixel and neighbors' dependence. Classical techniques like maximum likelihood classification does not account for this dependence, yielding unsatisfactory outcomes. This chapter also emphasizes the accuracy of multirresolution processes when applied to segmentation, the kind of processes inherent to the human vision system. In Chapter 3 a very thorough and detailed account on vector quantization can be found. This provides a model and a compression technique suitable for transmitting information through computing networks. Chapter 4 provides very refreshing material about supervised classification. 1-D and 2-D hidden Markov models (HMM) are introduced and an algorithm based on 2-D\ HMM is presented. The algorithm includes several steps to estimate the parameters and the well known EM-algorithm is used. Its computational complexity is also analysed. Real images are considered and the classification techniques are compared by showing to what extent the proposed algorithm and some other competitors manage to distinguish natural landmarks from some other objects. NEWLINENEWLINENEWLINEThe classification algorithm shows a remarkable performance. Wavelet transforms are used in the compression process to take out the most outstanding features of the images. In Chapter 5 2-D multirresolution HMM are defined. They allow for another way of understanding images as a set of matrices of decreasing dimension. Thereby, the original image corresponds to the highest resolution while the image with the lowest resolution is chosen according to some criteria. Several references are given in which classification can be enhanced by combining features obtained at different levels of resolution. Comparisons are made in a similar fashion to that exposed in Chapter 4, by considering real images. Chapter 6 presents some naive procedures for testing. Null hypothesis are restricted to the multivariate normal distribution and Markov model. Chapter 7 deals with an algorithm suitable for segmentation and compression altogether, based on 2-D HMM models. In the end, Chapter 8 extends the 2-HMM models to Markov models which allows for joint compression and restauration. Everyone interested in image processing can understand the main ideas in this book, which are clearly exposed and well stated. Anyway, mathematical details are also included when necessary to provide in-depth treatment in many cases.