Integrated Computer-Aided Engineering

This paper proposes a neural architecture, based on two Hopfield nets interconnected with a Boltzmann Machine, for a completely data driven edge-preserving restoration of blurred and noisy images. Solving this restoration problem entails the joint estimation of the image, the degradation operator and the noise statistics, assuming that only the data are available. Since we consider the class of piecewise smooth images, modeled through a coupled Markov Random Field with an explicit, constrained line process, the hyperparameters of the image model must be estimated as well. Adopting a fully Bayesian approach, the solution can be obtained by the joint maximization of a suitable distribution with respect to the image field, the model hyperparameters, and the degradation parameters. The very high computational complexity of this joint maximization means that in most practical cases it cannot be applied, unless some approximations are adopted. In this paper, by exploiting the presence of an explicit and binary line field, we propose some approximations which are effective in computing the solution by means of an architecture based on interacting neural networks. In particular, we propose an architecture where the main computational load is supported by two Hopfield nets, one computing the intensity field, the other performing a least square estimation of the blur coefficients. The Boltzmann Machine is used following two modalities: running and learning. In the running modality, it updates the binary line process; in the learning modality, it performs the ML estimation of the hyperparameters, which are interpreted as the weights of cliques of interconnected neurons. Simulation results are provided to highlight the feasibility and the efficiency of the adopted methodology.

There are two fundamental problems to be solved in any scalable computer system: tolerate and hide latency of remote accesses, and, tolerate and hide idling due to synchronization among parallel processes. Architectures which can not solve these issues will fail in building large-scale parallel processing systems. One possible solution for tolerating memory and synchronization latency is the introduction of threads and fast context switching mechanism among threads. Systems which support this technique are called multithreaded systems.

Multimedia applications usually require large computing power and thus, massivelly parallel systems are good candidates for such tasks. Additionally, multimedia applications usually involve the processing of huge amount of data (e.g. audio or video information), therefore both the classical shared or distributed memory parallel systems may be inadequate for fulfilling all the needs. Finally, multimedia applications (e.g. image processing) in some cases may require other computing model than current commodity RISC processors can provide.

A range of multithreaded architectures can be idealistic for multimedia applications, which is massively parallel, has distributed memory for the sake of scalability. Such architectures, which support remote memory accesses, may be a proper combination of different computing models, e.g. von Neumannn and dataflow ones.

In this paper, the design space of multithreaded architectures is introduced, and a certain architecture, called KUMP/D (Kyushu University Multimedia Processor on Datarol-II) is described. It is also shown how a multithreaded architecture can be built in a short design cycle by using a commercial high-end microprocessor and easily programmable hardware devices.

Different tasks in image processing exhibit different computational requirements that should be considered with respect to the architecture. This is particularly critical in parallel machines where many parallelization techniques, as data partitioning and mapping on processors, use of shared memory space, exploitation of pipelining with pre-fetching affect dramatically the performance with a strong relation with algorithm and architectural parameters.

The paper defines computational models for tightly-coupled multiprocessors with crossbar architecture, both for data-parallel local algorithms and for global algorithms such as spatial transformations. To solve the intrinsic memory limitations of low-cost, highly integrated systems, the paper proposes to extend the classical block processing model by analytically modeling also the case of multiple processing stages.

The models have been compared in detail and have been efficiently adopted for optimizing performance in block processing on crossbar multiprocessors for low-level computer vision applications.

The structure of Field Programmable Gate Arrays (FPGAs) naturally fits that of fine grain array algorithms. The paper investigates the geometrical and layout-related implementation problems of FPGA-based processor arrays. A general methodology for implementing parametrized array processors on FPGAs is presented. Then, a detailed layout algorithm is proposed. A new feature of the algorithm is the uniform treatment of inter- and intra-module nets that allows the layout of the basic processor to be optimized with respect to the critical path of the whole, arbitrarily large processor array. The approach is demonstrated on a massively parallel processor array for binary morphology.

Motion estimation in image sequences is a fundamental issue in many applications as for instance in artificial vision and three-dimensional scene reconstruction. Among all the existing techniques, we are particularly interested in methods using Gabor filters, which are known to furnish quality results but usually require intensive calculation. A new, fast energy-based method is presented, which combines in a direct manner the energetic responses of Gabor spatio-temporal filters organized in triads. An implementation of this technique on a general purpose Digital Signal Processor (DSP) board is described and the advantages compared with Very Large Scale Integration (VLSI) and parallel machine approaches. Our hardware implementation attains a reasonably fast output rate (several images/second) for a better resolution than in the most recent VLSI implementations. These results open interesting perspectives for real-time implementations (such as in mobile robotics) and for the obtention of higher-level results by combining different Gabor filter techniques (for motion, edge detection, texture analysis, etc.).

An interactive system for content-based retrieval in a pictorial database using color distribution features is discussed. Images are internally represented and indexed by a set of histograms reflecting homogeneous chromatic properties of image structure and also including color-induced geometric features. User query images are interactively composed through a graphic environment featuring color sketching, image sampling and relevance feedback. Queries are first translated into the internal representation and then matched against the database index using a modified multi-feature histogram intersection operator. Examples of system operation are provided, and experimental results are reported indicating that using multiple distributions instead of global color histogramming allows for a system output in better accordance to human expectation.