Rajeev Kumar Singh

Fuzzy Counter Propagation Network for Freehand Sketches-Based Image Retrieval

In this paper, we present Fuzzy Counter Propagation Network (FCPN) for Sketch-Based Image Retrieval (SBIR) with collection of freehand sketches; trademark and clip art, etc., using feature descriptors. FCPN is combination of Counter Propagation Network (CPN) and Fuzzy Learning (FL). We use features descriptor like Histogram of Gradient (HOG) for freehand sketches/images and these features are used to the training of FCPN. Flicker dataset containing 33 different shape categories, is used for training and testing. Different similarity measure functions are discussed and used similarity between query by nonexpert sketchers and database. We compare proposed FCPN method with other existing Feed-forward Networks (FFN) and Pattern Recognition Network (PRN). Experimental results show that FCPN methods outperform over networks.

An Efficient Contrast Enhancement Technique Based on Firefly Optimization

In the modern environment, digital image processing is a very vital area of research. It is a process in which an input image and output might be either any image or some characteristics. In image enhancement process, input image, therefore, results are better than given input image for any particular application or set of objectives. Traditional contrast enhancement technique results in lightning of image, so here Discrete Wavelet transform is applied on image and modify only Low–Low band. In this presented technique, for enhancement of given image having low contrast apply Brightness Preserving Dynamic Histogram Equalization (BPHDE), Discrete Wavelet Transform (DWT), Thresholding of sub-bands of DWT, Firefly Optimization and Singular Value Decomposition (SVD). DWT divides image into 4 bands of different frequency: High–high (HH), High–low (HL), Low–high (LH), and Low–low (LL). First apply a contrast enhancement technique named brightness preserving dynamic histogram equalization technique for enhancement of a given low-contrast image and boosts the illumination, then apply Firefly optimization on these 4 sub-bands and thresholding applied, this optimized LL band information and given input image’s LL band values are passed through SVD and new LL band obtained. Through inverse discrete wavelet transform of obtained new LL band and three given image’s HH, HL, and LH band obtained an image having high contrast. Quantitative metric and qualitative result of presented technique are evaluated and compared with other existing technique. A result reveals that presented technique is a more effective strategy for enhancement of image having low contrast. The technique presented by this study is simulated on Intel I3 64-bit processor using MATLAB R2013b.

Biogeography particle swarm optimization based counter propagation network for sketch based face recognition

In this paper, we present a Biogeography Particle Swarm Optimization (BPSO) based Counter Propagation Network (CPN) i.e. BPSO-CPN for Sketch Based Face Recognition (SBFR) system. A new criterion of selecting exemplar vector using biogeography learning based PSO is used for optimization of Mean Square Error (MSE) between feature vector of sketch and photo. In this work, we use Histogram of Gradient (HOG) feature vector for similarity measures between sketch and photo. Select a sketch as query image from database and using BPSO-CPN to retrieves similar photos from database. Proposed BPSO-CPN method is tested on CUHK and IIITD sketch dataset containing about 1000 sketches and photos. The experimental result envisage that, BPSO-CPN gives promising results and achieves high precision as comparison with other existing methods and neural networks. Motivation behind this research work is to find missing or wanted persons who involve in antinational activities and it help investigating agencies to narrow down the suspects quickly.

Kohonen neural network model reference for nonlinear discrete time systems

In this work, an adaptive neural network like Kohonen neural network (KNN) model reference is used for tracking control of nonlinear system. Proposed adaptive Kohonen neural network (ADKNN) are used to minimize the error between output and target signal for nonlinear discrete-time systems. The ADKNN is a feed-forward neural network help for approximation of the nonlinearities in the industrial plant and main characteristic of the system is taken into account is disturbances in the system. Tracking error by the adaptive ADKNN based approximation system is an important characteristic for the design and analysis. It is shown in results that the preference of the error system is decisive to the solution of tracking control. Difference between ADKNN output and reference signal can be made arbitrarily small in the close neighbourhood of zero. The viability of the ADKNN is verified via simulation example of nonlinear system.

ADAPTIVE NEURAL NETWORK FOR SKETCH BASED IMAGE RETRIEVAL

In this paper, we present neural network approach for Sketch Based Image Retrieval (SBIR) using Histogram of Gradient (HOG) feature descriptor. This paper emphasis on back propagation Feed-forward Network (FFN) and Pattern Recognition Network (PRN) used for sketch based retrieval. Neural network is a popular tool used for pattern recognition and approximation of unknown nonlinear functions. We use features descriptor like Histogram of Gradient (HOG) for free hand and human face sketches and these features are used to training of network. Experimental results and analysis are based on CHUK and Flicker dataset, used for training and testing. Different similarity measure functions are discussed and used similarity between query by non-expert sketchers and database.

Object extraction using topological models

Object extraction from complex scene of image is important for computer vision and pattern recognition. Fully automatic object extraction without human intervention is very tedious. Therefore, good solution of object extraction from complex scene is obtain by interactively. In this paper an interactive topological model based object extraction are proposed, which is based on two main important properties (i) initially low level segmentation and (ii) similarity measures between regions. Regions of input image are merged using Bhat-Nayar similarity. The extensive experiment has been performed to the efficiency and effectiveness of proposed method.