Theo E. Schouten

Human-centered object-based image retrieval

A new object-based image retrieval (OBIR) scheme is introduced. The images are analyzed using the recently developed, human-based 11 colors quantization scheme and the color correlogram. Their output served as input for the image segmentation algorithm: agglomerative merging, which is extended to color images. From the resulting coarse segments, boundaries are extracted by pixelwise classification, which are smoothed by erosion and dilation operators. The resulting features of the extracted shapes, completed the data for a -vector. Combined with the intersection distance measure, this vector is used for OBIR, as are its components. Although shape matching by itself provides good results, the complete vector outperforms its components, with up to 80% precision. Hence, a unique, excellently performing, fast, on human perception based, OBIR scheme is achieved.

Three-dimensional fast exact Euclidean distance (3D-FEED) maps

In image and video analysis, distance maps are frequently used. They provide the (Euclidean) distance (ED) of background pixels to the nearest object pixel. Recently, the Fast Exact Euclidean Distance (FEED) transformation was launched. In this paper, we present the three dimensional (3D) version of FEED. 3D-FEED is compared with four other methods for a wide range of 3D test images. 3D-FEED proved to be twice as fast as the fastest algorithm available. Moreover, it provides true exact EDs, where other algorithms only approximate the ED. This unique algorithm makes the difference, especially there where time and precision are of importance.

Timed Fast Exact Euclidean Distance (tFEED) maps

In image and video analysis, distance maps are frequently used. They provide the (Euclidean) distance (ED) of background pixels to the nearest object pixel. In a naive implementation, each object pixel feeds its (exact) ED to each background pixel; then the minimum of these values denotes the ED to the closest object. Recently, the Fast Exact Euclidean Distance (FEED) transformation was launched, which was up to 2x faster than the fastest algorithms available. In this paper, first additional improvements to the original FEED algorithm are discussed. Next, a timed version of FEED (tFEED) is presented, which generates distance maps for video sequences by merging partial maps. For each object in a video, a partial map can be calculated for different frames, where the partial map for fixed objects is only calculated once. In a newly developed, dynamic test-environment for robot navigation purposes, tFEED proved to be up to 7x faster than using FEED on each frame separately. It is up to 4x faster than the fastest ED algorithm available for video sequences and even 40% faster than generating city-block or chamfer distance maps for frames. Hence, tFEED is the first real time algorithm for generating exact ED maps of video sequences.

Performance prediction of large MIMD systems for parallel neural network simulations

Abstract In this paper, we present a performance prediction model for indicating the performance range of MIMD parallel processor systems for neural network simulations. The model expresses the total execution time of a simulation as a function of the execution times of a small number of kernel functions, which have to be measured on only one processor and one physical communication link. The functions depend on the type of neural network, its geometry, decomposition and the connection structure of the MIMD machine. Using the model, the execution time, speedup, scalability and efficiency of large MIMD systems can be predicted. The model is validated quantitatively by applying it to two popular neural networks, backpropagation and the Kohonen self-organizing feature map, decomposed on a GCel-512, a 512 transputer system. Measurements are taken from network simulations decomposed via dataset and network decomposition techniques. Agreement of the model with the measurements is within 1–14%. Estimates are given for the performances that can be expected for the new T9000 transputer systems. The presented method can also be used for other application areas such as image processing.

M-HinTS: Mimicking Humans in Texture Sorting

Various texture analysis algorithms have been developed the last decades. However, no computational model has arisen that mimics human texture perception adequately. In 2000, Payne, Hepplewhite, and Stoneham and in 2005, Van Rikxoort, Van den Broek, and Schouten achieved mappings between humans and artificial classifiers of respectively around 29% and 50%. In the current research, the work of Van Rikxoort et al. was replicated, using the newly developed, online card sorting experimentation platform M-HinTS: http://eidetic.ai.ru. nl/M-HinTS/. In two separate experiments, color and gray scale versions of 180 textures, drawn from the OuTex and VisTex texture databases were clustered by 34 subjects. The mutual agreement among these subjects was 51% and 52% for, respectively, the experiments with color and gray scale textures. The average agreement between the k-means algorithm and the participants was 36%, where k-means approximated some participants up to 60%. Since last years results were not replicated, an additional data analysis was developed, which uses the semantic labels available in the database. This analysis shows that semantics play an important role in human texture clustering and once more illustrate the complexity of texture recognition. The current findings, the introduction of M-HinTS, and the set of analyzes discussed, are the start of a next phase in unraveling human texture recognition.

Video Surveillance using Distance Maps

Human vigilance is limited; hence, automatic motion and distance detection is one of the central issues in video surveillance. Hereby, many aspects are of importance, this paper specially addresses: efficiency, achieving real-time performance, accuracy, and robustness against various noise factors. To obtain fully controlled test environments, an artificial development center for robot navigation is introduced in which several parameters can be set (e.g., number of objects, trajectories and type and amount of noise). In the videos, for each following frame, movement of stationary objects is detected and pixels of moving objects are located from which moving objects are identified in a robust way. An Exact Euclidean Distance Map (E2DM) is utilized to determine accurately the distances between moving and stationary objects. Together with the determined distances between moving objects and the detected movement of stationary objects, this provides the input for detecting unwanted situations in the scene. Further, each intelligent object (e.g., a robot), is provided with its E2DM, allowing the object to plan its course of action. Timing results are specified for each program block of the processing chain for 20 different setups. So, the current paper presents extensive, experimentally controlled research on real-time, accurate, and robust motion detection for video surveillance, using E2DMs, which makes it a unique approach.

A Neural Network Approach to Spectral Mixture Analysis

Imaging spectrometers acquire images in many narrow spectral bands. Because of the limited spatial resolution, often more than one ground cover category is present in a single pixel. In spectral mixture analysis the fractions of the ground cover categories present in a pixel are determined, assuming a linear mixture model. In this paper neural network methods which are able to perform this analysis are considered. Methods for the construction of training and test data sets for the neural network are given. Using data from 3 spectrometers with 6, 30 and 220 bands and 3 or 4 ground cover categories, it is shown that a back-propagation neural network with one hidden layer is able to learn the relation between the intensities of a pixel and its ground cover fractions. The distributions of the differences between true and calculated fractions show that a neural network performs the same or better than a conventional least squares with covariance matrix method. The calculation of the fractions by a neural network is much faster than by the least squares methods, training of the neural networks requires however a large amount of computer time.

Decomposition of mixed pixels

Within the instantaneous field of view of a scanning device often more than one object is included, resulting in a pixel in which several characteristics are mixed. Classically the proportions of the components of such a mixed pixel are estimated using a linear mixture model. In this paper a new method is introduced for estimating the characteristics of these components, from which their proportions can be derived. Experiments with simulated data sets are conducted to compare the methods regarding their accuracy on estimating the proportions. In addition it is determined how well the proposed method can estimate the characteristics of each component.

Incremental Distance Transforms (IDT)

A new generic scheme for incremental implementations of distance transforms (DT) is presented: Incremental Distance Transforms (IDT). This scheme is applied on the city-block, Chamfer, and three recent exact Euclidean DT (E2DT). A benchmark shows that for all five DT, the incremental implementation results in a significant speedup: 3.4 -10 times. However, significant differences (i.e., up to 12.5 times) among the DT remain present. The FEED transform, one of the recent E2DT, even showed to be faster than both city-block and Chamfer DT. So, through a very efficient incremental processing scheme for DT, a relief is found for E2DTs computational burden.

Distance transforms: academics versus industry

In image and video analysis, distance transformations (DT) are frequently used. They provide a distance image (DI) of background pixels to the nearest object pixel. DT touches upon the core of many applications; consequently, not only science but also industry has conducted a significant body of work in this field. However, in a vast majority of the cases this has not been published in major scientific outlets but has been filed as a patent application. This article provides a brief introduction into DT, including a specification of a few of the most prominent algorithms in the field. Next, a few interesting algorithms from the last decade are discussed. A benchmark including eight DT algorithms (i.e., city block, Danielsson’s algorithm, chamfer 3-4, hexadecagonal region growing, a recent claimed true Euclidean DT, and three exact Euclidean DT) has been executed, which illustrates the intriguing complexity of DT in terms of precision and computational complexity. Subsequently, a selection of key patent applications are discussed that have emerged in this field, including their scientific merit and areas of application. Finally, this article’s findings are summarized and discussed, with an emphasis on both the common ground of scientific articles and patent applications as well as the added value they can have to each other.