This paper proposes a fast method for image segmentation. After an optimal split of the image into rectangular regions, this paper focuses on the fast merging of these regions. Since the computation time is very small, hence it is suitable for real time applications, while producing a good segmentation for tracking purposes.

Micro-pipelines are linear (1D) structures for asynchronous communications. In retinotopic VLSI vision chips, communicating over 2D image regions is a key to efficient mid-level vision computations. However, micro-pipelines are limited to 1D communications only. In this paper, we introduce an extension of the micro-pipeline, called the convergent micro-pipeline, for implementing mixed asynchronous-synchronous regional computations over arbitrary shaped regions. This operator is exploited in programmable artificial retinas (PAR), that is, image sensors with a digital processor in each pixel, for low power vision applications. To illustrate the behavior and versatility, several regional computations are described.

Low level local image processing is efficiently performed by array processors operating in SIMD mode. Performing mid-level regional image processing leads to using local combinatorial operators combined with an asynchronous programmable interconnection network. However, this approach has an important hardware cost because asynchronism implies the use of combinatorial operators with many inputs. This cost should be reduced for a dense VLSI implementation. We propose to increase the connectivity level of the interconnection network as a means to use only 2-input asynchronous combinatorial operators. Results are presented on the example of the regional sum mid-level primitive in vision chips. An extension of the methodology to higher connectivity levels is then proposed.

We present a robust implementation of a motion detection algorithm based on a markovian relaxation both on General Purpose Processors, and on a specialized architecture, the Associative Mesh. The Mesh architecture is an instance of the associative nets model targeting real time execution of low level image algorithms and vision-Soc implementation. The algorithm implementation on both architectures is described.

Several hard problems have to be addressed in order to parallelize image analysis algorithms. Indeed, at the region level, these algorithms handle irregular (and sometimes strongly dynamic) data-structures. Moreover, they often lead to an unbalanced amount of computations, which is quite impossible to foresee offline. This paper focus on the parallelization of the ANET image analysis programming environment. Thanks to graph related data structures and efficient computing primitives, ANET allows rapid image algorithm prototyping. But in return, these primitives are difficult to parallelize. We present a solution for powerful implicit parallelization of the ANET environment, without any change in the application programming interface. The ANET API is summarized and illustrated with some examples. Several parallelization experimentations are reported. The solution we propose is detailed, and results are given on complete image analysis applications. ANET appears as a powerful environment, both for its expressiveness that allows rapid prototyping and for its implicit parallelization that allows good computation time.

The split and merge segmentation method is a well-known method to extract homogeneous regions from an image. It consists to recursively split the image in four regular quadrants until every segment is homogeneous, then to merge the obtained regions as long as the merging respects some homogeneity criterion. From the experiment results an efficient optimized HP split criterion is obtained by using scan-based implementation.

The paper presents a modified version of the classical split and merge algorithm (Horowitz, S. and Pavlidis, T., 1976). Instead of performing a regular decomposition of the image, it relies on a split at an optimal position that makes a good interregion separation. The implementation of the algorithm uses an initial image preprocessing to speed-up computation. Experimental results show that the number of regions generated by the split phase is largely reduced and that the distortion of the segmented image is smaller, while the execution time is slightly increased.

This paper presents the current evolution of the associative mesh project. It aims at the design of a reconfigurable, asynchronous and massively parallel SIMD architecture, targeted towards image analysis implementation. Its basic principle relies on the use of global operations (associations) that, given any interpixel connection graph, can compute global operations over connected sets of these graphs. One of our current objectives is the implementation of an associative mesh with a SoC-type circuit. In this paper, we examine which architectural modifications would this approach imply. We also consider the benefits brought by this technique and the repercussions on the design's performances

Many works have been done for parallelizing low-level image analysis computations. However the task is harder for higher levels, as the data manipulations are complex, and there is a wide range of algorithms to encompass. To allow concurently speed and programmability, a high-level programming model that can be efficiently implemented on parallel architectures is required. To achieve this goal, we propose the associative nets model, a parallel computing model for image analysis based on simple data-parallelism paradigms, providing special features, such as graph-based data structures to handle irregular data, virtual data-structures to ease hierarchical image descriptions, and specific primitives (dirassoc) to compute on the interpixels relation graph. For implementation purposes, the dirassoc computing primitive performs asynchronous local computations until it reaches stability. Asynchronism has many, advantages for hardware (speed, power consumptions, and chip size) as well as in software (less synchronization barriers). However to insure completion of the asynchronous operation, the dirassoc must use a set of specific operators (r-operators) introduced by Ducourthial. In this paper we emphasize on the interest of the r-operators and of the asynchronous computations for image analysis algorithms. We give applications in distance transforms, contour closing, Voronoi segmentation, watershed segmentation, and mathematical morphology. Hence, we show that asynchronous computations arc powerful tools for image analysis on interpixels graphs.

The quantization of coefficient transform with the orthogonal transformation and the filtering with a low pass filter is often used to reduce the psychophysical redundancy. The aim of this paper is to present the improvements of a method we developed for still image compression, namely the GDCT. The improvements of the GDCT by means of a nonlinear quantization and by selecting an optimal filtering relying on a statistical criterion. An alternative quantization based on quadtree decomposition is also discussed.

The proposed multiresolution fractal encode are images compression schemes that combine one of orthogonal transformation Dct. They improve the performance of conventional fractal compression algorithm: blocking effect and image blurring by better coding of high compression frequencies.

This paper presents some results of programming efficient matching algorithms on a new asynchronous parallel programming model. Matching algorithms are widely used in image processing when considering high-level treatments. Pattern analysis, database search, 2D and 3D reconstruction all need matching algorithms to perform. Experiments we did were mainly oriented towards a particular matching problem: the stable marriage algorithm. Different implementations of this algorithm have been done on a massively parallel asynchronous model. This model relies on a network of asynchronously communicating processors leading to very fast SIMD treatments. The asynchronous model and implementations of the matching algorithm are presented. An example of image processing problem is also used for illustration purpose and supports the architectural discussion and results

In this paper we present the programming environment Anet for image analysis, that aims to bridge the gap between programmability requirements and parallel efficiency. It is based on the graph based associative nets computing model, and allows irregular data manipulation. As it is intrinsically a parallel model, parallel execution can be quite naturally considered, and as the number of primitives is small, effective parallelization requires an initial limited effort and can be reused by a large set of programs.

With the recent developments in multimedia and telecommunication technologies, content-based information retrieval is becoming increasingly important support for various areas such as digital libraries, interactive video,... The algorithms involved in real multimedia application need to handle, as quickly as possible, a large volume of data that is a time consuming which can limit the impact of this application. In this paper we describe a system, based on a configurable processor, designed to accelerate the indexing document information (text, images, table...). The architecture is suitable for European STRETCH project (Storage and RETrieval by Content of imaged documents).

A content-based information retrieval is well used in many applications (digital libraries, interactive video, medical, ...). The methods involved in content-based information retrieval algorithms need to handle, as quickly as possible, a big volume of data. We are involving in European STRETCH project (Storage and RETrieval by Content of imaged documents) that deals with the archiving and the retrieval by content of imaged (scanned) documents. A component extraction is an important step in the archiving process and it is a time consuming. In this paper we describe a system, based on a configurable processor and a configurable network, designed to accelerate the extraction of the homogeneous components (text, images, table, etc.) of scanned documents.

A chip that is able to detect an equilibrium state in CMOs gates is designed. It allows one to know whether a system is in a steady state in each one of its components or if some output is still changing. This can be useful to detect the completion of an operation in the context of asynchronous circuits. Simulations and test results are given.

This paper presents a computing model named associative mesh and its application for image analysis. The model relies on global computations on subgraphs of an image, as a primitive operation to manipulate objects in an image. While the model is adapted to an efficient hardware implementation, it is also quite expressive in terms of image analysis. The paper illustrates that by presenting the implementation of Voronoï diagrams based image segmentation.

This paper presents a new parallel computing model called Associative Nets. This model relies on basic primitives called associations that consist of applying an associative operator over connected components of a subgraph of the physical interprocessor connection graph. Associations can be very efficiently implemented (in terms of hardware cost or processing time) thanks to asynchronous computation. This model is quite effective for image analysis and several other fields; as an example, graph processing algorithms are presented. While relying on a much simpler architecture, these algorithms have, in general, a complexity equivalent to the one obtained by more expensive computing models, like the PRAM model.