Nonlinear and Adaptive Techniques in Digital Image Processing, Analysis and Computer Vision

A multichannel nonlinear vector median based edge enhancing filter has been developed for colour image processing. The first filtering stage consists of three subfilters; whereas, the final output is obtained through comparing the subfilter outputs and the vector median output. The resulting multichannel filter is shown to enhance degraded edges in colour images. Nonlinear multivariate image filtering techniques are proposed to handle colour images corrupted by noise. Three reduced orderings are defined by selecting three different central locations. Considering noise attenuation, edge preservation and detail retention, multivariate filters are designed by combining R-ordering schemes. Locally adaptive versions of these techniques are also developed. Another locally adaptive multichannel filter based on the concept of trimmed mean filtering is developed for the purpose of colour image enhancement. The output of the filter is the mean of some subset selected from the input samples in the current window. These subsets are selected using the distance information from the marginal median and the centre sample. Suitable choices for the size of these subsets control the noise attenuation and detail preservation of the proposed filter. The next scheme developed for u image restoration decomposes the filtering process into two phases: directional estimation and magnitude estimation. The two phases can be computed simultaneously .The directional estimate is obtained by the basic directional filter; while, the magnitude estimate can be obtained by a number of scalar or vector filtering. Vector median (VM) and VM alternating with the mean filter have been used in a non-adaptive and adaptive schemes, respectively, yielding good results for colour image restoration. The final technique developed for colour image restoration concerns images which have been highly corrupted with impulse noise and for which conventional processing fails.

Nonlinear transforms for object recognition research focused on the development of invariants for grey scale images. From the extension of the theory of invariants to the group of motions in space, important new results arose. Invariant grey scale features are characteristics of grey scale images which remain constant if the images are transformed according to the action of a transformation group. Such features have several applications in computer vision and pattern. Algorithms for determining translation- and rotation invariant features from grey scale images. Have been developed. These features are calculated in two steps. First, a local nonlinear function has to be evaluated for every pixel of the image and afterwards the results of the local computations must be summed. This strategy can be used to determine the properties of the features for scenes with more than one object. It is explained how to construct features which are invariant even if the objects in the scene are rotated and translated independently. Moderate occlusions are tolerable. Furthermore it is shown how to use these techniques for the recognition of articulated objects. It has to be emphasized that the algorithms work directly with the grey values and do not rely on the extraction of geometric primitives like edges or corners in a preprocessing step. All algorithms have been implemented and tested both on synthetic and real image data. The algorithms have been implemented on a massively parallel system using the parallel image processing library developed in the project. An increase in the adaptivity of the features can be achieved by a parameterized calculation. Inspired by neural network, techniques have been developed a feedforward network. These techniques can be expected to have applications ,for example, in visual inspection and quality control.

A noisy image database has been developed and has been used in the performance comparisons by all the partners. In its present form, the database contains 42, 512x512 8-bit images obtained by corrupting two original images 'airfield' and 'bridge' with a contaminated Gaussian noise that serves as a basis for performance cross comparison of the algorithms. The algorithms developed for ultrasonic image filtering and analysis have successfully been applied to simulated ultrasound B-mode data and displayed ultrasound image data for speckle suppression and image segmentation. Morphological signal-adaptive median filters, optimal stack filters, weighted order statistics filters have successfully been applied to grey-scale images. The noisy database is used for testing the performance of filtering algorithms applied to grey-scale images. The performance of the algorithms serves as a baseline of the performance of any filtering algorithm applied to grey-scale images. The multichannel nonlinear filters for colour image restoration and colour image enhancement have been applied to still colour images and colour image sequences with very satisfactory results. The algorithms for moving object tracking have successfully been applied to real image sequences for tracking vehicles in normal road traffic scenes. Invariant grey-scale features have been applied for texture classification in visual inspection of fabrics. They have also been applied for building indexing functions for content-based image retrieval from video databases.

Ultrasonic imaging has become an important modality in the field of medical imaging systems. Ultrasonic images suffer from a special kind of noise called speckle. Signal-adaptive techniques for speckle noise removal have been developed. Both envelope-detected and displayed ultrasonic images have been considered. Motivated by the observation that for displayed ultrasonic data the maximum likelihood (ML) estimator of the original (noiseless) signal closely resembles the L2 mean (ie the ML estimator of the original signal for envelope-detected data), signal-adaptive L2 mean filters have been designed for both cases. The derived filters have a homomorphic structure (ie the L2 mean filter) due to the speckle noise statistics (Rayleigh distribution or signal-dependent Gaussian distribution). Secondly, the segmentation of ultrasonic images using self-organizing neural networks has been implemented. A novel variant of Learning Vector Quantizer (LVQ) neural network has been developed that is able to segment ultrasonic images in classes representing various tissue and lesion characteristics. This is the so called L2 LVQ algorithm. It can be combined with signal-adaptive filtering in order to allow preservation of image edges and details as well as maximum speckle reduction in homogeneous regions. The design of filtering processes combining segmentation and optimum L-filtering, and their use for the suppression of speckle noise in ultrasonic images has been proved very successful in practice.

The effort spent in the parallelization of nonlinear image processing algorithms resulted in the implementation of the general purpose parallel image processing system (PIPS). The system is designed for current parallel distributed memory systems and comprises implementations of some hundreds of image processing operations which are fully scalable and portable. Moreover, high efficiencies have been achieved within a broad range of applications. The system is ideally suited to speed up scientific and industrial image processing applications in terms of software development cycles and computational power. The great variety of parallel distributed memory systems require the use of several abstraction concepts for the system independent software design. The systems differ especially in the layout of the communication facilities between the processing elements. Message passing systems as an abstraction of the physical interconnection scheme offer some portability. In order to achieve scalable and efficient solutions one has to consider the a priori knowledge that is available and include it in the concrete realization of the software. This has been done for a great amount of linear and nonlinear image processing algorithms that form together with the required abstraction layers and infrastructure the general purpose PIPS. Thereby special emphasis was taken for flexible data distribution layout, the minimization of communication demands and data transfer times and the modularity of the system that enables the fast integration of new applications. The main application areas of the system are: iterative algorithms for image restoration (used for image restoration of the optical distortion of the Hubble Space Telescope); a tomographic analysis of the three-dimensional temperature distribution of approximately rotational symmetric flames; restoration of distorted images with additive noise; extraction of translational and rotational invariant features for object recognition purposes; image segmentation techniques based on watershed algorithms.

Fast design procedures for a subclass of stack filters, called 'weighted order statistic' (WOS) filters in the real domain was one of the major results that have been obtained. The optimization of WOS filters in the real domain domain saves an enormous amount of computation time as well as storage. Other adaptive structures for optimal design of WOS filters were also proposed. These include both locally adaptive and trainable algorithms. The different techniques developed can be implemented in very large scale integration (VLSI). Chip/boards implementing the algorithms may figure in new television sets, video recorder/players, cameras and other imaging devices. The algorithms developed were shown to outperform existing algorithms in the literature through analytic derivation and computer simulations. Other areas that might benefit from the results developed include medicine, oil exploration and mining, computer vision (particularly fault inspection and on-line tasks).

A new Radial Basis Functions neural network whose training is based on marginal median has been developed which improves motion representation and modelling for dynamic scene understanding. Motion representation and modelling is an important step towards dynamic image understanding An algorithm for simultaneous estimation and segmentation of the optical flow has been implemented. In this approach, the moving scene is decomposed in different regions with respect to their motion, by means of a Radial Basis Functions (RBF) Neural Network. The learning algorithm for the RBF network employs the median algorithm for estimating the centres of the Radial Basis Functions and the Median of Absolute Deviations (MAD) for estimating their variances. The proposed algorithm has been applied in real image sequences and has been compared to other competitive algorithms (eg. the Iterated Conditional Modes). It has been found that it outperforms the other algorithms.

A parallel implementation of an existing watershed algorithm has been developed using a single program-multiple data (SPMD) approach. Additional analysis and development of existing and new watershed algorithms, based on a split-and-merge approach, have been conducted under this activity. The resulting parallel watershed segmentation algorithm is based on sequential scanning. It is a fast SPMD algorithm which computes the watershed image by integrating the morphological gradient. The algorithm is well suited for single instruction multiple data (SIMD) computers since no ordered queues are needed. A faster SPMD algorithm has been developed for image segmentation based on rain falling simulation. Good segmentation results are efficiently obtained by computing the steepest slope lines of a topographic surface on a distributed memory computer with message passing communication. The different techniques developed can be implemented in very large scale integration (VLSI). Chips/boards implementing the algorithms may figure in new television sets, video recorder/player, cameras and other imaging devices. Other areas that might benefit from the results developed include medicine, oil exploration and mining, computer vision (particularly fault inspection and on-line tasks).

Noise corruption of images is a frequently encountered problem in many image processing tasks. The need emerges for implementing smoothing techniques that are able to treat different kinds of noise. The main objectives of image filtering algorithms are: the smoothness of noise in homogeneous regions; the preservation of edges; the removal of impulses of constant as well as of random value. A class of filters that fulfils these requirements is the so called signal-adaptive filters. The morphological signal-adaptive median filter is a paradigm of this class. It performs well on many kinds of noise and does not require the a priori knowledge of a noise-free image, but only of certain noise characteristics easily estimated. It adapts its behaviour based on a local signal to noise (SNR) measurement achieving thus edge preservation and noise smoothing in homogeneous regions. Through the use of an anisotropic local window adaptation procedure, by employing binary morphological erosions and dilations with predefined structuring elements, the largest possible window size is determined at each pixel location thus allowing better noise smoothing in flat regions or edge borders without edge blurring. In the filter capabilities, satisfactory smoothness of impulsive noise is included due to its enhanced impulse detection mechanism able to further detect randomly-valued impulses. The filter proves to perform very well in severe noise corruption cases.

Adaptive stack filtering and synthesis resulted in a simple approach to the design of stack filters under the mean absolute error criterion. The estimates of the Boolean function output decisions were derived from the probabilities of each possible binary window sequence. These probabilities were obtained by comparing corrupted and uncorrupted versions of the same image, or parts of an image. It was found that, in most cases, this simplified method produced identical results to those obtained from the method incorporating a check for the stacking property. The method was tested and the filter produced reasonable results. Fast and efficient algorithms for stack filtering were developed. In addition, a detailed analysis of the behaviour of certain subclasses of stack filters has been conducted. The major results include a full training framework which has been set for optimal Boolean and stack filtering, making optimal design a manageable task. An application of a stack related filter as a code-division multiple-access (CDMA) detector was developed using the matched median filter. In this application, a matched median is used as a detector for direct-sequence multiple-access communication. Its performance is evaluated in a synchronous CDMA environment, where two types of interference are assumed: multiple-access interference and additive impulsive channel noise. Average bit-error probabilities computed using Monte-Carlo simulations have shown the effectiveness of the proposed scheme. The different techniques developed can be implemented in very large scale integration (VLSI). Chip/boards implementing the algorithms may figure in new television sets, video recorder/players, cameras and other imaging devices. The algorithms developed were shown to outperform existing algorithms in the literature through analytic derivation and computer simulations. Other areas that might benefit from the results include medicine, oil exploration and mining, computer vision (particularly fault inspection and on-line tasks).

Exploitable results

Share this page

Download