Dimo Dimov

Towards a Multinational Car License Plate Recognition System

A full-fledged image-based car license plate recognition (CLPR) system is described in the paper. CLPR provides an inexpensive automatic solution for remote vehicle identification. Gray-level input images are assumed. The localization stage of the CLPR yields a plate clip followed by character segmentation and recognition. The recognition scheme combines adaptive iterative thresholding with a template-matching algorithm. The method is invariant to illumination and is robust to character size and thickness, skew and small character breaks. Promising results have been obtained in the experiments with Israeli and Bulgarian license plates including images of poor quality. Also, the possibility of using an “off-the-shelf” OCR has been explored.

Graph cuts optimization for multi-limb human segmentation in depth maps

We present a generic framework for object segmentation using depth maps based on Random Forest and Graph-cuts theory, and apply it to the segmentation of human limbs in depth maps. First, from a set of random depth features, Random Forest is used to infer a set of label probabilities for each data sample. This vector of probabilities is used as unary term in α-β swap Graph-cuts algorithm. Moreover, depth of spatio-temporal neighboring data points are used as boundary potentials. Results on a new multi-label human depth data set show high performance in terms of segmentation overlapping of the novel methodology compared to classical approaches.

Adaptive license plate image extraction

The paper represents the automatic plate localization component of a Car License Plate Recognition system. The approach concerns stages of preprocessing, edge detection, filtering, detection of the plates position, slope evaluation, and character segmentation and recognition. Single frame gray-level images are used as the only source of information. In the experiments Israeli and Bulgarian license plates were used, camera obtained at different daytime and whether conditions. The results derived have shown that the approach is robust to illumination, plate slope, scale, and is insensitive to plates country peculiarities. These results could be also usable for other applications in the input-output transport systems, where automatic recognition of registration plates, shields, signs, etc., is often necessary.

Human limb segmentation in depth maps based on spatio-temporal Graph-cuts optimization

We present a framework for object segmentation using depth maps based on Random Forest and Graph-cuts theory, and apply it to the segmentation of human limbs. First, from a set of random depth features, Random Forest is used to infer a set of label probabilities for each data sample. This vector of probabilities is used as unary term in α-β swap Graph-cuts algorithm. Moreover, depth values of spatio-temporal neighboring data points are used as boundary potentials. Results on a new multi-label human depth data set show high performance in terms of segmentation overlapping of the novel methodology compared to classical approaches.

CBIR approach to the recognition of a sign language alphabet

The task of recognizing letters from the sign language alphabet, by means of which the hearing impaired people finger-spell words and proper nouns, is interpreted as a CBIR (Content Based Image Retrieval) problem. An arbitrary input image of given sign (palm gesture) is treated as a sample for search within a database (DB), which contains a large enough set of images (i.e. projections from a sufficient number of view points) for each letter of the sign language alphabet. We assume that the gestures for recognition are static images, which have been appropriately extracted from the input video sequence. In addition, we have at our disposal a CBIR method for image DB access that is simultaneously fast enough and noise-tolerant. The paper describes both the methodology used for building up the DB of image samples and the experimental study for the noise-tolerance of the available CBIR method. The latter is used to acknowledge the applicability of the proposed approach.

Fast, shape based image retrieval

The paper presents a method for fast access by content into database (DB) of images. The method proposed is based on the image tree of contours defined for graphic images, as well as on the one-dimensional complex Fourier transform of the contours. In this way, problems connected with the image normalization concerning translation, rotation, scaling, reflection and intensity are currently solved. The most informative data of the image are ordered by importance in a key of fixed length, on which the fast access is performed using the well-known index access methods of a conventional DB management system (DBMS). The method is tested on a DB of about 4000 images of hallmarks. The experimental system EIRS, intended for the Bulgarian Patent Office practice, is also described here in brief.

Cyclic histogram thresholding and multithresholding

The paper concerns the problem of thresholding of an integer domain of 1D cyclic histogram (periodic function) resulting in two or more consecutive regions (classes). An optimal solution is searched for in the terms of the statistical criterion well known in the pattern recognition area as Fishers LDA (Linear Discriminant Analysis) and also successfully applied for image binarization by Otsu (1979). An effective (quadratic complexity) extension of the Otsus method is also known, which segments the image by respective thresholding of the image intensity histogram into arbitrary number of classes. We propose one more extension of this approach for the case of the cyclic histograms. Similar problem can be brought by the optimal segmentation of color images based on their HSV histogram, and more general in all problems which try to approximate a given periodic function with a predefined number of Gaussians. The paper describes the theoretical basis and the experimental evaluation of the proposed approach.

Wavelet transform application to fast search by content in database of images

The paper presents a fast access method to images in a conventional DB (DataBase). The method is based on two-dimensional wavelet transform of images preliminary normalized by size, orientation and intensity. The image content for search is considered the normalized image graphics that should be well localizable into the input query picture. The most essential image data are represented as a key of fixed length, on which the fast access is performed using the index access methods of a conventional DB management system. The method is experimented on a DB of about 4000 images of trademarks.

Real time video stabilization for handheld devices

The paper proposes a method and robust algorithm for 2D video stabilization designed for portable devices in real-time. The BSC (Boundary Signal Computation) chip of TI (Texas Instruments) is essentially used (or emulated herein) for searching of correlations between the 1D integral projections, horizontal and vertical ones, by a SAD (Sums of Absolute Differences) approach. The proposed method is based on an accurate vector model allowing interpretations of increasing complexity for the transformations among frames. Experiments, conducted on testing video clips, are very promising for the future R&D of the method.

Appearance-Based 3D Object Approach to Human Ears Recognition

The paper presents an approach for recognition of 3D objects using a database (DB) of precedents. Each object of interest for recognition is presented in the DB through a sufficient number of 2D projections (images), each from a different view point. If it is available a CBIR method to access the DB that is to be fast enough and noise resistant, the number of necessary view positions for each 3D object can be substantially reduced, for example, to several tens or a few hundreds of images. The authors have already applied successfully this appearance-based approach two times: i) for recognition of palm signs from a sign language alphabet and ii) for human face recognition. The recent advance in 3D scanning technologies allow to fresh up the training phase of the proposed method, i.e. the DB gathering of the necessary appearance of precedents for 3D objects, now more accurately and simply. At the same time, the true recognition remains based on images from conventional 2D cameras. This study aims to experiment the mentioned approach for the case of human ears recognition, which, according to our research is of interest to the guild on Biometrics, in the country, in Europe and worldwide.