A computer vision system is developed for 3-D object recognition using artificial neural networks and a model-based top-down feedback analysis approach. This system can adequately address the problems caused by an incomplete edge map provided by a low-level processor for 3-D representation and recognition. The system uses key patterns that are selected using a priority assignment. The highest priority is given to the key pattern with the most connected node and associated features. The features are space invariant structures and sets of orientation for edge primitives. The labeled key features are provided as input to an artificial neural network for matching with model key patterns. A Hopfield-Tank network is appiled to two levels of matching to increase the computational effectiveness. The first matching is to choose the class of the possible model and the second matching is to find the model closest to the candidate. The result of such matchings is utilized in generating the model-driven top-down feedback analysis. This model is then rotated in 3-D space to find the best match with the candidate and to provide the additional features in 3-D. In the case of multiple objects, a dynamic search strategy is adopted to recognize objects using one
pattern at a time. This strategy is also useful in recognizing occluded objects. The experimental results are presented to show the capabiilty and effectiveness of the system.
We propose a new approach for detecting ellipses. The approach is based on the geometrical property that five points on an ellipse can determine the parameters of the ellipse, and the symmetry of ellipses is used to obtain these points. Using symmetry, we classify the edge points in an input image into several subimages. Ellipses with different symmetric axes will lie in different subimages. In each subimage, symmetry is applied again to obtain those sets of five points that possibly lie on the same ellipse. Finally, using geometrical properties and the Hough transform, we extract all ellipses in an input image successfully. The proposed method can detect partially occluded ellipses. Experimental results show that the proposed method is fast and robust.
We present a parallel mechanism for detecting contours embedded in a binary image. The proposed algorithm can detect contours in parallel in sequential machines with less computational effort. The hardware architecture of the algorithm is also proposed. Experiments with a wide variety of binary images show that the speed of this new technique is much faster than that of other contour detection metho
The error diffusion algorithm is a flexible and powerful procedure to quantize pictures and holograms. Its free parameters allow controlled transformation of the quantization noise. The influence of the object on the quantization noise is shown. A method is presented to modify the object before it is quantized to reduce the object-dependent component of the quantization noise.
An efficient algorithm for color image quantization is proposed based on a new vector quantization technique that we call sequential scalar quantization. The scalar components of the 3-D color vector are individually quantized in a predetermined sequence. With this technique, the color palette is designed very efficiently, while pixel mapping is performed with no computation. To obtain an optimal allocation of quantization levels along each color coordinate, we appeal to the asymptotic quantization theory, where the number of quantization levels is assumed to be very large. We modify this theory to suit our application, where the number of quantization 1evels is typically small. To utilize the properties of the human visual system (HVS), the quantization is performed in a luminance-chrominance color space. A luminance-chrominance weighting is introduced to account for the greater sensitivity of the HVS to luminance than to chrominance errors. A spatial activity measure is also incorporated to reflect the increased sensitivity of the HVS to quantization errors in smooth image regions. The algorithm yields high-quality images and is significantly faster than existing quantization algorithms.
We introduce new morphological filters, called soft morphological filters. They maintain most of the desirable properties of standard morphological operations yet are less sensitive to additive noise and to small variations in the shapes of the objects to be filtered. The main difference from standard morphological filters is that maximum and minimum operations are replaced by more general weighted-order statistics. This results in the loss of some algebraic properties but improved performance under noisy conditions.
The demand for fast and cost-effective access to multiple compressed data sources is imminent. An interactive link between a distant user and the compressed data is needed such that data transmission is constrained to only a small subset of the compressed data that possesses specific features of interest to the user, hence avoiding transmission and decompression of other noninteresting data sources. We present the first step toward developing this user-compressed data link. We show that oriented line features can be detected in data that are transformed, prior to being quantized and coded, using the discrete cosine transform (DCT). Our work is based on the DCT performed on block sizes of 32 x 32 pixels. The choice of this particular block size is due to proprietary constraints imposed by the specific problem we were commissioned to solve. It involves a proprietary database structure. The DCT is the basis of many established image compression standards such as the Joint Photographic Experts Group (JPEG) and the Motion Picture Experts Group (MPEG). These standards are based on block sizes of 8 x 8 pixels. We also show that the effect we report is exploitable with this more familiar block size. Theoretical proof of the DCT line-feature detector and experimental results are provided. An extension of the approach for logical inference of the existence of specific objects of interest is also outlined. When used as a feature analytical tool in conjunction with its compression role, the DCT may serve as a smart compression procedure for intelligent data archiving, abstraction, and transmission.
A simple yet efficient image data compression method is presented. This method is based on coding only those segments of the image that are perceptually significant to the reconstruction of the image. Sequences of image pixels whose gray-level differences from the pixels of the previous row exceed two prespecified thresholds are considered significant. These pixels are coded using a differential pulse code modulation scheme that uses a 15-level recursively indexed nonuniform quantizer for the first pixel in a segment and a 7-level recursively indexed nonuniform quantizer for all other pixels in the segment. The quantizer outputs are Huffman coded. Simulation results show that this scheme can obtain subjectively satisfactory reconstructed images at low bit rates. It is also computationally very simple, which makes it amenable to fast implementation.
The blue noise mask (BNM) is a halftone screen that produces unstructured, visually pleasing halftone images. Since it is a point process, halftoning using the BNM can be implemented considerably faster than error diffusion and other algorithms. However, in the construction of the original BNM, a number of constraints were used to limit its characteristics in the spatial and frequency domains. These constraints were not efficient to compute and required adaptability to all gray levels in the construction process. The original BNM also contained some small but unwanted low-frequency components at some gray levels. In this paper, we present a revised approach to the generation of blue noise patterns and the construction of BNMs employing more efficient computations and eliminating more unwanted residual low-frequency components. Psychovisual evaluation shows that dithering with the new BNM gives excellent results and its rating is statistically indistinguishable from that of error diffusion with serpentine raster and perturbed weights.
The toner jet method has been previously proposed to perform electrophotographic nonimpact printing easily. To clarify the fundamental properties of this method, the jet behavior of toner is studied through simulation on a personal computer. The mesh electrode is assumed to be inserted halfway between the development
roller and the paper-back electrode. This is done to control the locus and the distribution path of toner from the magnetic development roller electrode to the electrode on the back of the paper receiver. The electric field applied between the magnetic development roller and the paper-back electrode is higher than the field between the mesh electrode and the magnetic development roller. The locus and the distribution of toner particles developed on the paper are simulated by changing the applied voltage in each row and each column of the mesh electrode. It is assumed that the jet behavior of toner particles from the magnetic development roller to paper is controllable. In conclusion, the useful role of the mesh electrode in the image quality of the toner jet method is suggested.