Real-time image registration is potentially an enabling technology for the effective and efficient use of many image-guided diagnostic and treatment procedures relying on multimodality image fusion or serial image comparison. Mutual information is currently the best-known image similarity measure for intensity-based multimodality image registration. The calculation of mutual information is memory intensive and does not benefit from cache-based memory architectures in standard software implementations, i.e., the calculation incurs a large number of cache misses. Previous attempts to perform image registration in real time focused on parallel supercomputer implementations, which achieved significant speedups using large, expensive supercomputers that are impractical for clinical deployment. We present a hardware architecture that can be used to accelerate a number of linear and elastic image registration algorithms that use mutual information as an image similarity measure. A proof-of-concept implementation of the architecture achieved speedups of 30 for linear registration and 100 for elastic registration against a 3.2 GHz Pentium III Xeon workstation. Further speedup can be achieved by using several modules in parallel.

Due to consumers' demand for faster picture shot time in the rapidfy expanding digital still camera market, it is of importance to address the real-time implementation issues in the development of passive automatic focusing for digital still cameras. This article discusses such real-time implementation issues that are often overlooked when designing passive contrast sensing automatic focusing on digital still camera processors. Specifically, algorithmic design tradeoffs between automatic focusing speed, accuracy, and power consumption, are addressed. A sample implementation and its performance results on an actual digital still camera hardware platform powered by the Texas Instruments TMS320DM270 processor are presented to further convey these real-time implementation issues.

In this article, we present the BinDCT algorithm, a fast approximation of the Discrete Cosine Transform, and its efficient VLSI architectures for hardware implementations. The design objective is to meet the real-time constrain in embedded systems. Two VLSI architectures are proposed. The first architecture is targeted for low complexity applications such as videophones, digital cameras, and digital camcorders. The second architecture is designed for high perform applications, which include high definition TV and digital cinema. In order to meet the real-time constrain for these applications, we decompose the structure of the BinDCT algorithm into simple matrices and map them into multi-stage pipeline architectures. For low complexity implementation, the proposed 2-D BinDCT architecture can be realized with the cost of 10 integer adders, 80 registers and 384 bytes of embedded memory. The high performance architecture can be implemented with an extra of 30 adders. These designs can calculate real-time DCT/IDCT for video applications of CIF format at 5 MHz clock rate with 1.55 volt power supply. With its high performance and low power consumption features, BinDCT coprocessor is an excellent candidate for real-time DCT-based image and video processing applications.

This article presents a system for acquisition, filtering, and segmentation of thermographic images in real time. Image acquisition is carried out using an infrared line scanner (IRLS), with which thermographic line scans are captured from hot strips while they are moving forward along a track. During the acquisition process, a relationship between each sample in the line scan and its position on the strip is established using a theoretical model of the IRLS, whose parameters have been adjusted using a calibration procedure. After the acquisition, line scans are filtered using a new signal operator designed to work in real time. Online with acquisition and filtering processes, segmentation is applied to the stream of line scans to group them into regions with similar temperature pattern. Two new segmentation algorithms based on well-known approaches, region merging and edge detection, have been designed to work in real time on a stream of line scans. The algorithms are evaluated using a novel segmentation assessment method based on the uncertainty of the ground truth, which can also be used for parameter tuning. Experimental results from a database of 200,000 images taken from manufactured steel strips over a period of three years demonstrate the efficiency and effectiveness of the proposed system.

Imaging systems are traditionally developed using structured analysis and design techniques. While there are many reasons that engineers choose this approach, one is the expected real-time performance benefits. But structured approaches tend to be rigid with respect to changing needs, technologies, devices, and algorithms. More generally, these systems are difficult or impossible to reuse because each new problem requires a new solution. Object-oriented approaches, on the other hand, can lead to systems that are more readily reused if certain best practices are followed. However, the conventional wisdom is that the price for such benefits is degraded real-time performance. The contribution of this work is an examination of these best practices, in the form of patterns and design principles, with reference to imaging systems. Then an extensive implementation of these practices is done on an existing imaging system, Kahindu, which is a teaching package built using the object-oriented paradigm. We then show how by applying these best practices not only improved structure is obtained, but surprisingly, improved performance as well. Our results challenge the conventional belief that the “price” for the improved structure, ease-of-extension, maintainability, etc. of object-oriented systems in imaging systems is degraded performance.

We propose an intersection model and strategy for automatic road extraction from aerial imagery. The proposed approach is able to detect typical intersections such as crossroads, T-junctions and Y-junctions based on matching the model to the image features. Compared to the traditional morphological methods, for example the combination with thinning and 8-neighbor pattern matching, our approach is less sensitive to noise and holes, and less like to produce a false match. The road network is constructed by connecting the detected intersections. The connecting hypothesis is generated and validated using the road tracking method and the road shape including the width is refined using ribbon snakes. We show the feasibility of our approach by presenting results for a suburban area, and evaluating them in comparison to the existing road map.

The printability of paper is extremely dependent on the topographical properties of the substrate. Imaging instruments make it possible to obtain detailed 3D scans of paper surfaces that can be further used to calculate valuable quality predictors. A new imaging instrument, OptiTopo, based on the photometric stereo principle was developed at the Swedish Pulp and Paper Research Institute (STFI) with the advantages of an extreme acquisition speed and the possibility of simultaneously acquiring topographic and reflectance information. The topographical imaging of paper surfaces using this technique has now been investigated and improved. Eleven paper samples covering a wide range of different grades have been analyzed by OptiTopo and their scans compared to those obtained using a reference imaging technique. By applying a suitable signal treatment it is possible to improve the instrument's performance in terms of detail rendering capability. The positive results have been confirmed using visual assessment, classical statistical indicators and frequency analysis. The present limitations of the technique in relation to the physical properties of the substrate are discussed and absolute boundaries for the instrument are proposed.

The color rendition of a halftone image depends on, among other things, physical and optical dot gains and fluorescence of substrates. A unified model describing spectral reflectance of a print is proposed with consideration of these effects. In this model the effects of either physical or optical dot gain are characterized by a single parameter, while those of fluorescence by two sets of spectral parameters, one for fluorescence of bare paper and one for fluorescence of a print solid. This model is tested and further illustrated with applications to images generated by a laser color printer on ordinary office papers.

Color is widely used for content-based image retrieval. In these applications the color properties of an image are characterized by the probability distribution of the colors in the image. These probability distributions are very often estimated by histograms although the histograms have many drawbacks compared to other estimators such as kernel density methods. In this article we investigate whether using kernel density estimators instead of histograms could give better retrieval results based on hue descriptors of color images. In this article we introduce the Fourier series coefficients as descriptors of hue distributions. We argue that under certain conditions these coefficients are optimal in a least squared error sense. We will also apply Parseval formula to compute the similarity of two distributions directly from these Fourier coefficients. Our experiments show that this modification of the kernel based similarity estimation has better retrieval performance than the histogram methods and we will also show that the method is insensitive to parameter changes as long as they are selected in a reasonable range.