Recognition of handwritten digits is most challenging sub task of character recognition due to various shapes, sizes, large variation in writing styles from person to person and also similarity in shapes of different digits. This paper presents a robust Telugu language handwritten digit recognition system. The Telugu language is most popular and one of classical languages of India. This language is spoken by more than 80 million people. The proposed method initially performs preprocessing on input digit pattern for removing noise, slat correction, size normalization and thinning. This paper divides the preprocessed Telugu handwritten digits into four differential zones of 2x2, 3x3, 4x4 and 6x6 pixels and extracts 65 features using Fractal dimension (FD) from each zone. The proposed zonal fractal dimension (ZFD) method uses, Feed forward backward propagation neural network (FFBPNN) for classifying the digits with learning rate of 0.01 and sigmoid function as an activation function on extracted 65 features. This paper evaluated the efficiency of the proposed method based on 5000 Telugu handwritten digit samples, each consists of ten digits from different groups of people and totally 50,000 samples. The performance of classification of the proposed method also evaluated using statistical parameters like recall, precision, F-measure and accuracy.

We present a new technique for content based image retrieval by deriving a Local motif pattern (LMP) code co-occurrence matrix (LMP-CM). This paper divides the image into 2 x 2 grids. On each 2 x 2 grid two different Peano scan motif (PSM) indexes are derived, one is initiated from top left most pixel and the other is initiated from bottom right most pixel. From these two different PSM indexes, this paper derived a unique LMP code for each 2 x 2 grid, ranges from 0 to 35. Each PSM minimizes the local gradient while traversing the 2 x 2 grid. A co-occurrence matrix is derived on LMP code and Grey level co-occurrence features are derived for efficient image retrieval. This paper is an extension of our previous MMCM approach [54]. Experimental results on popular databases reveal an improvement in retrieval rate than existing methods.

In this paper, two extended versions of motif co-occurrence matrices (MCM) are derived and concatenated for efficient content-based image retrieval (CBIR). This paper divides the image into 2 x 2 grids. Each 2 x 2 grid is replaced with two different Peano scan motif (PSM) indexes, one is initiated from top left most pixel and the other is initiated from bottom right most pixel. This transforms the entire image into two different images and co-occurrence matrices are derived on these two transformed images:  the first one is named as “motif co-occurrence matrix initiated from top left most pixel (MCMTL)” and second one is named as “motif co-occurrence matrix initiated from bottom right most pixel (MCMBR)”. The proposed method concatenates the feature vectors of MCMTL and MCMBR and derives multi motif co-occurrence matrix (MMCM) features. This paper carried out investigation on image databases i.e. Corel-1k, Corel-10k, MIT-VisTex, Brodtaz, and CMU-PIE and the results are compared with other well-known CBIR methods. The results indicate the efficacy of the proposed MMCM than the other methods and especially on MCM [19] method.

Skew detection and correction is one of the major preprocessing steps in the document analysis and understanding. In this paper we are proposing a new method called “Principle-axis farthest pairs Quadrilateral (PFPQ)”  mainly for detecting skew in the Telugu language document and also in other Indian languages. One of the popular and classical languages of India is Telugu language. The Telugu language is spoken by more than 80 million people. The Telugu language consists of simple and complex characters attached with some extra marks known as “maatras” and “vatthulu”. This makes the process of skewing of Telugu document is more complex when compared to other languages. The PFPQ, initially performs pre-processing and divides the text in to connected components and estimates principle axis furthest pair quadrilateral then removes the small and large portions of quadrilaterals of connected components. Then by using painting and directional smearing algorithms the PFPQ estimates the skew angle and performs the de-skew. We tested extensively the proposed algorithm with five different kinds of documents collected from various categories i.e., Newspapers, Magazines, Textbooks, handwritten documents, Social media and documents of other Indian languages. The images of these documents also contain complex categories like scientific formulas, statistical tables, trigonometric functions, images, etc. and encouraging results are obtained. 

One of the popular descriptor for texture classification is the local binary pattern (LBP). LBP and its variants derives local texture features effectively. This paper integrates the significant local features derived from uniform LBPs(ULBP) and threshold based conversion factor non-uniform (NULBP) with complete textons. This integrated approach represents the complete local structural features of the image.   The ULBPs are proposed to overcome the wide histograms of LBP. The ULBP contains fundamental aspects of local features.  The LBP is more prone to noise and this may transform ULBP into NULBP and this degrades the overall classification rate. To addresses this, this paper initially transforms back, the ULBPs that are converted in to NULBPs due to noise using a threshold based conversion factor and derives noise resistant fundamental texture (NRFT) image. In the literature texton co-occurrence matrix(TCM) and multi texton histogram (MTH) are derived on a 2x2 window. The main disadvantage of the above texton groups is they fail in representing complete textons. In this paper we have integrated our earlier approach “complete texton matrix (CTM)” [16] on NRFT images. This paper computes the gray level co-occurrence matrix (GLCM) features on the proposed NRFCTM (noise resistant fundamental complete texton matrix) and the features are given to machine learning classifiers for a precise classification. The proposed method is tested on the popular databases of texture classification and classification results are compared with existing methods.

This paper presents a complete image feature representation, based on texton theory proposed by Julesz’s, called as a complete texton matrix (CTM)for texture image classification. The present descriptor can be viewed as an improved version of texton co-occurrence matrix (TCM) [1] and Multi-texton histogram (MTH) [2]. It is specially designed for natural image analysis and can achieve higher classification rate. TheCTM can express the spatial correlation of textons and can be considered as a generalized visual attribute descriptor. This paper initially quantized the original textures into 256 colors and computed color gradient from RGB vector space. Then the statistical information of eleven derived textons, on a 2 x 2 grid in a non-overlapped manner are computed to describe image features more precisely. To reduce the dimensionality the present paper extended the concept of present descriptor and derived a compact CTM (CCTM). The proposed CTM and CCTM methods are extensively tested on the Brodtaz, Outex and UIUC natural images. The results demonstrate the superiority of the present descriptor over the state-of-art representative schemes such as uniform LBP (ULBP), local ternary pattern (LTP), complete –LBP (CLBP), TCM and MTH.

Texture classification and analysis are the most significant research topics in computer vision. Local binary pattern (LBP) derives distinctive features of textures. The robustness of LBP against gray-scale and monotonic variations and computational advantage have made it popular in various texture analysis applications. The histogram techniques based on LBP is complex task. Later uniform local binary pattern’s (ULBP) are derived on LBP based on bit wise transitions. The ULBP’s are rotationally invariant. The ULBP approach treated all non-uniform local binary pattern’s (NULBP) into one miscellaneous label. This paper presents a new texture classification method incorporating the properties of ULBP and grey-level co-occurrence matrix (GLCM). This paper derives ternary patterns on the ULBP and divides the 3 x 3 neighborhood in to dual neighborhood. The ternary pattern mitigates the noise problems particularly near uniform regions. The dual neighborhood reduces the range of texture unit from 0 to 6561 to 0 to 80. The GLCM features extracted from ULBP-dual texture matrix (ULBP-DTM) provide complete texture information about the image and reduce the texture unit range. Various machine learning classifiers are used for classification purpose. The performance of the proposed method is tested on Brodtaz, Outex and UIUC’s textures and compared with GLCM, texture spectrum (TS) and cross-diagonal texture matrix (CDTM) approaches.

The local binary pattern (LBP) and local ternary pattern (LTP) are basically gray scale invariant, and they encode the binary/ ternary relationship between the neighboring pixels and central pixel based on their grey level differences and derives a unique code. These traditional local patterns ignore the directional information. The proposed method encodes the relationship between the central pixel and two of its neighboring pixel located in different angles (α, β) with different directions. To estimate the directional patterns, the present paper derived variation in local direction patterns in between the two derivates of first order and derived a unique First order –Local Direction variation pattern (FO-LDVP) code. The FO-LDVP evaluated the possible direction variation pattern for central pixel by measuring the first order derivate relationship among the horizontal and vertical neighbors (0o Vs.90o; 90o Vs. 180o ; 180o Vs.270o ; 270o Vs. 0o) and derived a unique code. The performance of the proposed method is compared with LBP, LTP, LBPv, TS and CDTM using the benchmark texture databases viz. Brodtaz and MIT VisTex. The performance analysis shows the efficiency of the proposed method over the existing methods. 

Authorship attribution is one of the important problem, with many applications of practical use in the real-world. Authorship identification determines the likelihood of a piece of writing produced by a particular author by examining the other writings of that author. Most of the research in this field is carried out by using instance based model. One of the disadvantages of this model is that it treats the different documents of each author differently. It produces a matrix per each document of the author, thus creating a huge number of matrices per author, i.e. the dimensionality is very high. This paper presents authorship identification using Author based Rank Vector Coordinates (ARVC) model. The advantage of the proposed ARVC model is that it integrates all the author's profile documents into a single integrated profile document (IPD) and thus overcomes the above disadvantage. To overcome the ambiguity created by common words of authors ARVC model removes the common words based on a threshold. Singular value decomposition (SVD) is used on IPD after removing the common words. To reduce the overall dimension of the matrix, without affecting its semantic meaning a rank-based vector coordinates are derived. The eigenvector features are derived on ARVC model. The present paper used cosine similarity measure for author attribution and carries out authorship attribution on English poems and editorial documents

Face recognition is one of the important and popular visual recognition problem due to its challenging nature and its diverse set of applications. That's why face recognition is attracted by many researchers. Methods based on Local Binary Pattern (LBP) are widely used for face recognition in the literature, and it is sensitive to noise. To address this present paper utilized the powerful local texture descriptor that is less sensitive to noise and more discriminant in uniform regions called as Local Ternary Pattern (LTP). The Uniform Local Binary Pattern (ULBP) derived on LBP treats a large set of LBP under one label called as miscellaneous. This may result some loss of information on LBP and LTP based methods. To address this two Prominent LBP (PLBP) are derived, namely PLBP-Low (L) and PLBP-High (H) on LTP. Based on this the present paper derived eight texture features on facial images. A distance function is used on proposed texture features for effective face recognition. To eliminate most of the effects of illumination changes that are present in human face an efficient preprocessing method is used that preserves the significant appearance details that are needed for face recognition. The present method is experimented on Yale, Indian and American Telephone and Telegraph Company (AT&T) Olivetti Research Laboratory (ORL) data bases and it has given state-of-the-art performance on the three popular datasets.

Texture image retrieval plays a significant and important role in these days, especially in the era of big-data. The big-data is mainly represented by unstructured data like images, videos and messages etc. Efficient methods of image retrieval that reduces the complexity of the existing methods is need for the big-data era. The present paper proposes a new method of texture retrieval based on local binary pattern (LBP) approach. One of the main disadvantages of LBP is, it generates 256 different patterns on a 3x3 neighborhood and a method based on this for retrieval needs 256 comparisons which is very tedious and complex. The retrieval methods based on uniform LBP's which consists of 59 different patterns of LBP is also complex in nature. To overcome this, the present paper divided LBP into dual LBP's consisting four pixels. The present paper based on this dual LBP derived a 2-dimensional dual uniform LBP matrix (DULBPM) that contains only four entries. The texture image retrieval is performed using these four entries of DULBPM. The proposed method is evaluated on the animal fur, car, leaf and rubber textures. 

Extraction of flower regions from complex background is a difficult task, it is an important part of flower image retrieval, and recognition .Image segmentation denotes a process of partitioning an image into distinct regions. A large variety of different segmentation approaches for images have been developed. Image segmentation plays an important role in image analysis. According to several authors, segmentation terminates when the observer's goal is satisfied. For this reason, a unique method that can be applied to all possible cases does not yet exist. This paper studies the flower image segmentation in complex background. Based on the visual characteristics differences of the flower and the surrounding objects, the flower from different backgrounds are separated into a single set of flower image pixels. The segmentation methodology on flower images consists of five steps. Firstly, the original image of RGB space is transformed into Lab color space. In the second step 'a' component of Lab color space is extracted. Then segmentation by two-dimension OTSU of automatic threshold in 'a-channel' is performed. Based on the color segmentation result, and the texture differences between the background image and the required object, we extract the object by the gray level co-occurrence matrix for texture segmentation. The GLCMs essentially represent the joint probability of occurrence of grey-levels for pixels with a given spatial relationship in a defined region. Finally, the segmentation result is corrected by mathematical morphology methods. The algorithm was tested on plague image database and the results prove to be satisfactory. The algorithm was also tested on medical images for nucleus segmentation.