Bulusu Lakshmana Deekshatulu

Texture element feature characterizations for CBIR

Colour and texture are the most common features used in CBIR systems today. In this paper, we wish to investigate structural methods of texture analysis for CBIR in view of their closeness to human perception and description of texture. In structural analysis, local patterns are the key (as is the case with humans), and when used as features may be expected to return more relevant images in CBIR. One method to describe local patterns in computationally simple terms is texture spectrum proposed by He and Wang. In this paper, we propose two additional characterizations of local patterns. The first is an extension of He and Wangs texture spectrum to larger and more meaningful windows, along with new structural features that capture local patterns such as horizontal and vertical stripes, alternating dark and bright spots, etc. The second is a new method that characterizes patterns as contrast variations in 5 /spl times/ 5 windows. We apply the new texture characterizations to develop a CBIR application and tested their performance on two databases containing remote sensing images. Our results show accuracies that range from 60% to 100% depending on the query image and the features contained therein. These results indicate that our texture features are useful in retrieving images appropriate for different remote sensing applications.

A modified Gabor function for content based image retrieval

As the Gabor filters are direction dependent, the Gabor transform of an image is to be performed for all chosen directions. Thus the set of angles used in Gabor feature extraction does affect the results in applications such as Content Based Image Retrieval (CBIR). In the present work, we modify the Gabor filter suitably in such a way that the modified function besides being free from the choice of angles is as effective as the Gabor function itself. Additionally, our method of extraction of features is invariant to rotation in images. Our simulation results demonstrate that the modified Gabor based method being useful for CBIR shows better retrieval performance than the standard Gabor based method.

A wavelet based multiresolution algorithm for rotation invariant feature extraction

The present work aims at proposing a new wavelet representation formula for rotation invariant feature extraction. The algorithm is a multilevel representation formula involving no wavelet decomposition in standard sense. Using the radial symmetry property, that comes inherently in the new representation formula, we generate the feature vectors that are shown to be rotation invariant. We show that, using a hybrid data mining technique, the algorithm can be used for rotation invariant content based image retrieval (CBIR). The proposed rotation invariant retrieval algorithm, suitable for both texture and nontexture images, avoids missing any relevant images but may retrieve some other images which are not very relevant. We show that the higher precision can however be achieved by pruning out irrelevant images.

Neural network algorithm for parameter identification of dynamical systems involving time delays

This paper deals with the problem of identification of unknown parameters and time-delay of dynamical systems. A polynomial function is employed to modify the structure of the system and subsequently suitable algorithms for identifying the parameters of time-lag systems are developed using neural networks. Illustrative examples for identification of time-lag systems and quasi-linear systems with lag have been presented.

Multisensor Data Fusion Using Neural Networks

This paper presents a Hebbian learning based linear single-layer neural network based measurement fusion of multisensor data. The performance of the proposed unsupervised neural network algorithm is compared with traditional fusion methods based on Kalman filtering such as measurement fusion and state vector fusion. The experiments have been carried out using multisensor data obtained from different radars. The results demonstrate the viability of the proposed algorithm.

Towards Improving the Accuracy of Telugu OCR Systems

Design of a high accuracy OCR system is a challenging task as the system performance is affected by its component modules. Each module has its own impact on the overall accuracy of the OCR system. An improvement in a module reflects upon overall system performance. In the present work, we have developed an OCR system for Telugu. Our experiments on a corpus of about 1000 images has shown that the system performance is degraded due to broken characters caused by the binarization module as well as due to improper character segmentation. Therefore, we address the issues of handling broken characters and poor segmentation. A novel approach which is based on feedback from the distance measure used by the classifier is proposed to handle broken characters. For character segmentation, our proposed approach exploits the orthographic properties of Telugu script. As a result, significant improvement is obtained in the performance of the system. These algorithms are generic and may be applicable to other Indian scripts, especially to south Indian scripts. In our experiments, an end-to-end system performance is evaluated which is not reported in the literature.

A Novel Stroke Width Based Binarization Method to Handle Closely Spaced Thick Characters

Signboards and billboards provide a challenge to image segmentation methods, since these images may also have pictures and graphical objects, apart from text objects. Methods that often succeed in more traditional text block segmentation situations do not perform well here since estimation of text lines and character widths etc fail due to the short sample sizes. Further, extraction of characters of different font sizes, which can be found in the real world and signboard images, remains a problem. In this paper, as a solution to the mentioned problem, we propose two stroke width based binarization approaches. These approaches can be used to eliminate extraneous objects based upon estimates of stroke width. We compare our methods with several other stroke width based binarization methods. We observe that the previous approaches fail, when there are closely spaced thick characters. We show that our second approach is able to extract closely spaced thick characters better than any of the other methods.

A FOURIER-RADIAL DESCRIPTOR ALGORITHM FOR INVARIANT FEATURE EXTRACTION

By integrating the Fourier techniques and the edge information obtained using the radial symmetric functions, we propose in this paper an invariant feature extraction algorithm. Unlike the Gabor feature extraction method, the present method does not use direction dependent filters, nor does it use the images in polar form, for rotation invariance. Besides, the present Fourier-Radial invariant feature extraction algorithm, suitable for both the texture and non-texture images, has functional analogy with the Gabor feature extraction method, and hence, is easily implementable. It is mathematically proved, and justified through computations, that the method can generate the invariant and discriminative feature vectors. Our simulation results demonstrate that the method can be used for such applications as content-based image retrieval.

State Estimation and Tracking Problems: A Comparison Between Kalman Filter and Recurrent Neural Networks

The aim of this paper is to demonstrate the suitability of recurrent neural networks (RNN) for state estimation and tracking problems that are traditionally solved using Kalman Filters (KF). This paper details a simulation study in which the performance of a basic discrete time KF is compared with that of an equivalent neural filter built using an RNN. Real time recurrent learning (RTRL) algorithm is used to train the RNN. The neural network is found to provide comparable performance to that of the KF in both the state estimation and tracking problems. The relative merits and demerits of KF vs RNN are discussed with respect to computational complexity, ease of training and real time issues.

Spectral colour image processing

Colour is represented as Spectral Power Distribution (SPD) in physical 3D world scene. A set of three discrete values, triplet, such as RGB represents colour in digital 2D image. RGB values are derived from SPD using colour transforms. The physical basis of digital colour is SPD. Colour processing in natural 3D world scene could be thought of manipulation of SPD. We explored different colour enhancement techniques based on derived SPD. Examples showing the spectral based enhancement operations negative, flipping of wavelengths, morphological operations, changing illuminants, increasing the brightness, smoothing, edge detection etc. with experimental results on real images. The results appear to demonstrate that our proposed approach successfully simulates physical processes. Further, the approach outlined in this paper leads to numerous significant applications in computational photography.