- Open Access
Multi technique amalgamation for enhanced information identification with content based image data
© Das et al. 2015
- Received: 12 September 2015
- Accepted: 5 November 2015
- Published: 1 December 2015
Image data has emerged as a resourceful foundation for information with proliferation of image capturing devices and social media. Diverse applications of images in areas including biomedicine, military, commerce, education have resulted in huge image repositories. Semantically analogous images can be fruitfully recognized by means of content based image identification. However, the success of the technique has been largely dependent on extraction of robust feature vectors from the image content. The paper has introduced three different techniques of content based feature extraction based on image binarization, image transform and morphological operator respectively. The techniques were tested with four public datasets namely, Wang Dataset, Oliva Torralba (OT Scene) Dataset, Corel Dataset and Caltech Dataset. The multi technique feature extraction process was further integrated for decision fusion of image identification to boost up the recognition rate. Classification result with the proposed technique has shown an average increase of 14.5 % in Precision compared to the existing techniques and the retrieval result with the introduced technique has shown an average increase of 6.54 % in Precision over state-of-the art techniques.
- Image classification
- Image retrieval
- Otsu’s threshold
- Slant transform
- Morphological operator
- t test
Recent years have witnessed the digital photo-capture devices as a ubiquity for the common mass (Raventós et al. 2015). The low cost storage, increasing computer power and ever accessible internet have kindled the popularity of digital image acquisition. Efficient indexing and identification of image data from these huge image repositories has nurtured new research challenges in computer vision and machine learning (Madireddy et al. 2014). Automatic derivation of sematically-meaningful information from image content has become imperative as the traditional text based annotation technique has revealed severe limitations to fetch information from the gigantic image datasets (Walia et al. 2014). Conventional techniques of image recognition were based on text or keywords based mapping of images which had limited image information. It was dependent on the perception and vocabulary of the person performing the annotation. The manual process was highly time consuming and slow in nature. The aforesaid limitations have been effectively handled with content based image identification which has been exercised as an effective alternative to the customary text based process (Wang et al. 2013). The competence of the content based image identification technique has been dependent on the extraction of robust feature vectors. Diverse low level features namely, color, shape, texture etc. have constituted the process of feature extraction. However, an image comprises of number of features which can hardly be defined by a single feature extraction technique (Walia et al. 2014). Therefore, three different techniques of feature extraction namely, feature extraction with image transform, feature extraction with image morphology and feature extraction with image binarization have been proposed in this paper to leverage fusion of multi-technique feature extraction. The recognition decision of three different techniques was further integrated by means of Z score normalization to create hybrid architecture for content based image identification. The main contribution of the paper has been to propose fusion architecture for content based image recognition with novel techniques of feature extraction for enhanced recognition rate.
Reducing the dimension of feature vectors.
Successfully implementing fusion based method of content based image identification.
Statistical validation of research results.
Comparison of research results with state-of-the art techniques.
Three different techniques of feature extraction using image binarization, image transforms and morphological operators have been combined to develop fusion based architecture for content based image classification and retrieval. Hence, it is in correlation with research on binarization based feature extraction, transform based feature extraction and morphology based feature extraction from images. It is also in connection with research on multi technique fusion for content based image identification. Therefore, the following four subsections have reviewed some contemporary and earlier works on these four topics.
Feature extraction using image transform
Change of domain of the image elements has been carried out by using image transformation to represent the image by a set of energy spectrum. An image can be represented as series of basis images which can be formed by extrapolating the image into a series of basis functions (Annadurai and Shanmugalakshmi 2011). The basis images have been populated by using orthogonal unitary matrices as image transformation operator. This image transformation from one representation to another has advantages in two aspects. An image can be expanded in the form of a series of waveforms with the use of image transforms. The transformation process has been helpful to differentiate the critical components of image patterns and in making them directly accessible for analysis. Moreover, the transformed image data has a compact structure useful for efficient storage and transmission. The aforesaid properties of image transforms facilitate radical reduction of feature vector dimension to be extracted from the images. Diverse techniques of feature extraction has been proposed by exploiting the properties of image transforms to extract features from images using fractional energy coefficient (Kekre and Thepade 2009; Kekre et al. 2010). The techniques have considered seven image transforms and fifteen fractional coefficients sets for efficient feature extraction. Original images were divided into subbands by using multiple scales Biorthogonal wavelet transform and the subband coefficients were used as features for image classification (Prakash et al. 2013). The feature spaces were reduced by applying Isomap-Hysime random anisotropic transform for classification of high dimensional data (Luo et al. 2013).
Image binarization techniques for feature extraction
Feature extraction from images has been largely carried out by means of image binarization. Appropriate threshold selection has been imperative for execution of efficient image binarization. Nevertheless, various factors including uneven illumination, inadequate contrast etc. can have adverse effect on threshold computation (Valizadeh et al. 2009). Contemporary literatures on image binarization techniques have categorized three different techniques for threshold selection namely, mean threshold selection, local threshold selection and global threshold selection to deal with the unfavourable influences on threshold selection. Enhanced classification results have been comprehended by feature extraction from mean threshold and multilevel mean threshold based binarized images (Kekre et al. 2013; Thepade et al. 2013a, b). Eventually, it has been identified that selection of mean threshold has not dealt with the standard deviation of the gray values and has concentrated only on the average which has prevented the feature extraction techniques to take advantage of the spread of data to distinguish distinct features. Therefore, image signature extraction was carried out with local threshold selection and global threshold selection for binarization, as the techniques were based on calculation of both mean and standard deviation of the gray values (Liu 2013; Yanli and Zhenxing 2012; Ramírez-Ortegón and Rojas 2010; Otsu 1979; Shaikh et al. 2013; Thepade et al. 2014a).
Use of morphological operators for feature extraction
Commercial viability of shape feature extraction has been well highlighted by systems like Image Content (Flickner et al. 1995), PicToSeek (Gevers and Smeulders 2000). Two different categorization of shape descriptors namely, contour-based and region-based descriptors have been elaborated in the existing literatures (Mehtre et al. 1997; Zhang and Lu 2004). Emphasize of the contour based descriptors has been on boundary lines. Popular contour-based descriptors have embraced Fourier descriptor (Zhang and Lu 2003), curvature scale space (Mokhtarian and Mackworth 1992), and chain codes (Dubois and Glanz 1986). Feature extraction from complex shapes has been well carried out by means of region-based descriptors, since the feature extraction has been performed from whole area of object (Kim and Kim 2000).
Fusion methodologies and multi technique feature extraction
Information recognition with image data has utilized the features extracted by means of diverse extraction techniques to harmonize each other for enhanced identification rate. Recent studies in information fusion have categorized the methodologies typically into four classes, namely, early fusion, late fusion, hybrid fusion and intermediate fusion. Early fusion combines the features of different techniques and produces it as a single input to the learner. The process inherently increases the size of feature vector as the concentrated features easily correspond to higher dimensions. Late fusion applies separate learner to each feature extraction technique and fuses the decision with a combiner. Although it offers scalability in comparison to early fusion, still, it cannot explore the feature level correlations, since it has to make local decisions primarily. Hybrid fusion makes a mix of the two above mentioned techniques. Intermediate fusion integrates multiple features by considering a joint model for decision to yield superior prediction accuracy (Zhu and Shyu 2015). Color and texture features were extracted by means of 3 D color histogram and Gabor filters for fusion based image identification. The space complexity of the feature was further reduced by using genetic algorithm which has also obtained the optimum boundaries of numerical intervals. The process has enhanced semantic retrieval by introducing feature selection technique to reduce memory consumption and to decrease retrieval process complexity (ElAlami 2011). Local descriptors based on color and texture was calculated from Color moments and moments on Gabor filter responses. Gradient vector flow fields were calculated to capture shape information in terms of edge images. The shape features were finally depicted by invariant moments. The retrieval decisions with the features were fused for enhanced retrieval performance (Hiremath and Pujari 2007). Feature vectors comprising of color histogram and texture features based on a co-occurrence matrix were extracted from HSV color space to facilitate image retrieval (Yue et al. 2011). Visually significant point features chosen from images by means of fuzzy set theoretic approach. Computation of some invariant color features from these points was performed to gauge the similarity between images (Banerjee et al. 2009). Recognition process was boosted up by combining color layout descriptor and Gabor texture descriptor as image signatures (Jalab 2011). Multi view features comprising of color, texture and spatial structure descriptors have contributed for increased retrieval rate (Shen and Wu 2013). Wavelet packets and Eigen values of Gabor filters were extracted as feature vectors by the authors in (Irtaza et al. 2013) for neural network architecture of image identification. The back propagation neural network was trained on sub repository of images generated from the main image repository and utilizes the right neighbourhood of the query image. This kind of training was aimed to insure correct semantic retrieval in response to query images. Higher retrieval results have been apprehended with intra-class and inter-class feature extraction from images (Rahimi and Moghaddam 2013). In (ElAlami 2014), extraction of color and texture features through color co-occurrence matrix (CCM) and difference between pixels of scan pattern (DBPSP) has been demonstrated and an artificial neural network (ANN) based classifier was designed. In (Subrahmanyam et al. 2013), content-based image retrieval was carried out by integrating the modified color motif co-occurrence matrix (MCMCM) and difference between the pixels of a scan pattern (DBPSP) features with equal weights. Fusion of semantic retrieval results obtained by capturing colour, shape and texture with the color moment (CMs), angular radial transform descriptor and edge histogram descriptor (EHD) features respectively had outclassed the Precision values of individual techniques (Walia et al. 2014). Six semantics of local edge bins for EHD were considered which included the vertical and the horizontal edge (0,0), 45° edge and 135° edge of sub-image (0,0), non directional edge of sub-image (0,0) and vertical edge of sub-image at (0,1). Color histogram and spatial orientation tree has been used for unique feature extraction from images for retrieval purpose (Subrahmanyam et al. 2012).
Three different techniques of feature extraction have been introduced in this work namely, feature extraction with image binarization, feature extraction with image transform and feature extraction with morphological operator. However, there are popular feature extraction techniques like GIST descriptor which has much greater feature dimension compared to the proposed techniques in the work. GIST creates 32 feature maps of same size by convolving the image with 32 Gabor filters at 4 scales, 8 orientations (Douze et al. 2009). It averages the feature values of each region by dividing each feature map into 16 regions. Finally, it concatenates the 16 average value of all 32 feature maps resulting in 16 × 32 = 512 GIST descriptor. On the other hand, our approach has generated a feature dimension of 6 from each of the binarization and morphological technique. Feature extraction by applying image transform has yielded a feature size of 36. On the whole, the feature size for the fusion based classifier was (6 + 36 + 6 = 48) which is far less than GIST and has much lesser computational overhead. Furthermore, fusion based architecture for classification and retrieval have been proposed for enhanced identification rate of image data. Each of the techniques of feature extraction as well as the methods for fusion based architecture of classification and retrieval has been discussed in the following four subsections and the description of datasets has been given in the fifth subsection.
Feature extraction with image binarization
Binarization of the test images was carried out using the Otsu’s local threshold selection method. The process has been repeated for all the three color components to generate bag of words model (BoW) of features. Conventional BoW model has been based on SIFT algorithm which has a descriptor dimension of 128 (Zhao et al. 2015). Therefore, for three color components the dimension of the descriptor would have been 128 × 3 = 384. The size for SIFT descriptor has been huge and it has predestined problem for information losses and omissions as it has been found suitable only for the stability of image feature point extraction and description. Furthermore, the generated SIFT descriptors has to be clustered by k means clustering which has been based on allocation of cluster members by means of comparing squared Euclidian distance. The clustering process has been helpful to generate codewords for codebook generation which has been the final step of BoW. Process of k means clustering has huge computational overhead for calculating the squared Euclidian distance which eventually slows down the BoW generation. Hence, in our approach, the grey values higher than the threshold was clustered in higher intensity group and the grey values lower than the cluster was clustered in the lower intensity group. The mean of the two groups were calculated to formulate the codewords of higher intensity feature vectors and the lower intensity feature vectors respectively. Thus, each color component of a test image has been mapped to two codewords of higher intensity and lower intensity respectively. This has generated of codebook of size (3 × 2 = 6) for each image.
The algorithm for feature extraction has been stated in Algorithm 1 as follows:
Feature extraction using image transform
The algorithm for feature extraction using slant transform has been given in Algorithm 2.
Here the features were extracted in the form of visual words. Visual words have been defined as a small patch of image which can carry significant image information. The energy compaction property of Slant transform has condensed noteworthy image information in a block of 12 elements for an image of dimension (256 × 256). Thus, the feature vector extracted with slant transform was of size 12 for each color component which has given the dimension of feature vector as 36 (12 × 3 = 36) for three color components in each test image.
Feature extraction with morphological operator
Apply the gray scale opening operation to an image.
Peak = original image—opened image.
Display the peak.
The algorithm for feature extraction using morphological operator has been given in Algorithm 3.
Determination of image similarity measures was performed by evaluating distance between set of image features. Higher similarity has been characterized by shorter distance (Dunham 2009). A fusion based classifier, an artificial neural network (ANN) classifier and a support vector machine (SVM) classifier was used for the purpose. Each of the classifier types has been discussed in the following sections:
Fusion based classifier
Artificial neural network (ANN) classifier
The back propagation technique of multi layer perceptron has a significant role in supervised learning procedure. The network has been trained for optimization of classification performance by using the procedure of back propagation. For each training tuple, the weights were modified so as to minimize the mean squared error between the network prediction and the target value. These modifications have been made in the backward direction through each hidden layer down to the first hidden layer. The input feature vectors have been fed to the input units which comprised the input layer. The number of input units has been dependent on the summation of the number of attributes in the feature vector dataset and the bias node. The subsequent layer has been the hidden layer whose number of nodes has to be determined by considering the half of the summation of the number of classes and the number of attributes per class. The inputs that have passed the input layer have to be weighted and fed simultaneously to the hidden layer for further processing. Weighted output of the hidden layer was used as input to the final layer which has been named as the output layer. The number of units in the output layer has been denoted by the number of class labels. The feed forward property of this architecture does not allow the weights to cycle back to the input units.
Support vector machine (SVM) classifier
SVM has searched for the maximum separating hyperplane as shown in Fig. 6. The support vectors have been shown with thicker borders.
The algorithm was implemented using sequential minimal optimization (SMO) (Keerthi et al. 2001). The operating principle of SMO has been to select two Lagrange multipliers as the multipliers must obey a linear equality constraint. The two selected Lagrange multipliers jointly optimize to find the optimal value for these multipliers and updates the SVM to reflect the new optimal values.
Four different datasets namely Wang dataset, Oliva and Torralba (OT-Scene) dataset, Corel dataset and Caltech Dataset was used for the content based image recognition purpose. Each of the datasets has been described in the following subsections.
Oliva and torralba (OT-Scene) dataset
It was observed that classification with 0.024 % of the transform coefficient has the highest F1 Score and lowest MR compared to the rest. Hence, it was considered as the feature vector with a dimension of 36.
Precision and recall values for four public datasets using three feature extraction techniques
Feature extraction with binarization
Feature extraction with fractional coefficients of slant transform
Feature extraction with morphological operator
The comparison in Fig. 12 has clearly revealed that fusion based classification has shown an enhanced precision of 0.12, 0.13 and 0.067 compared to classification with ANN classifier for feature extraction with image binarization, partial transform coefficients and morphological operator respectively. The recall rate for classification with fusion based classification was also higher by 0.134, 0.141 and 0.08 in comparison to classification with ANN classifier for feature extraction with three above mentioned techniques.
It was observed that the proposed method has outclassed the existing techniques. It has an increased precision rate of 0.012, 0.108, 0.109, 0.178 and 0.228 and an enhanced recall rate of 0.037, 0.125, 0.126, 0.195 and 0.245 compared to the existing techniques, namely, (Thepade et al. 2014b; Yanli and Zhenxing 2012; Ramírez-Ortegón and Rojas 2010; Liu 2013; Shaikh et al. 2013) respectively as in Fig. 14. The proposed fusion technique was observed to have the maximum precision and recall values compared to the recent techniques cited in the literature.
The figure has clearly divulged that fusion technique of retrieval with classified query has fetched all the images of the same category to that of the query image, whereas, retrieval with generic or unclassified query has three images from classes other than the class of query in position 2, 15 and 19 respectively.
A comparison of retrieval with individual techniques of feature extraction and fusion based retrieval with classified query has been given in Fig. 15.
Results in Fig. 15 have shown an increase of 26.3, 34.5 and 19.5 % in precision values and enhancement of 5.26, 6.9 and 3.9 % in recall values for the fusion based retrieval technique with classified query in comparison to retrieval with individual feature extraction techniques. It was clearly established that the fusion based technique has outperformed the individual techniques.
Hypothesis 1: There is no significant difference among the Precision values of fusion based retrieval with classified query with respect to individual retrieval techniques
Statistical validation with paired t test
Retrieval by feature extraction with image transform
Retrieval by feature extraction with image binarization
Retrieval by feature extraction with morphological operator
The p values have clearly indicated significant difference in precision values of the fusion based retrieval technique with classified query compared to the existing techniques of retrieval. Hence, the null hypothesis was rejected and the proposed fusion technique with classified query has been found to boost the precision values with statistical significance.
The comparison in Fig. 17 has clearly established the superiority of the proposed fusion based retrieval technique with respect to existing fusion based technique of retrieval. The proposed retrieval technique has improved precision of 1.98, 3.2, 3.3, 3.49, 17.8, 21.1 and 26.31 % and superior recall of 0.4, 0.64, 0.66, 0.7, 3.56, 4.22 and 5.26 % compared to the existing fusion based techniques mentioned in Fig. 13.
The comparison shown in Fig. 18 has revealed an enhanced precision rate of 0.2, 0.5 and 2.1 % and increased recall rate of 0.04, 0.1 and 0.6 % respectively for the proposed method with respect to the existing semantic retrieval techniques.
It has reduced the dimension of feature vectors.
It has successfully implemented fusion based method of content based image identification.
The research results have shown statistical significance.
The research results have outperformed the results of state-of-the art techniques.
In depth analysis of feature extraction techniques have been exercised in this research work. Three different techniques of feature extraction comprising of image binarization, fractional coefficients of image transforms and morphological operations has been implemented to extract features from the images. The extracted features with multiple techniques were used for fusion based identification process. The proposed method of fusion has divulged statistical significance with respect to the individual techniques. The retrieval technique was implemented with classification as a precursor. The classification technique was used to classify the query image for retrieval. The method has shown better performance compared to generic query based method of retrieval. Thus, the importance of classification was established in limiting the computational overhead for content based image identification. Finally, image identification with the proposed technique has surpassed the state-of-the art methods for content based image recognition. The work may be extended towards content based image recognition in the field of military, media, medical science, journalism, e commerce and many more.
RD and ST have designed the feature extraction techniques and the classification and retrieval techniques. RD and SG have planned the statistical test and conclusion. RD wrote the manuscript. All the authors have read and approved the final manuscript.
The authors acknowledge Late Dr. H.B. Kekre for encouraging the experimental process. The authors also acknowledge Dr. Rohit Vishal Kumar and Dr. Subhajit Bhattacharya for explaining the statistical techniques.
The authors declare that they have no competing interests.
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
- Alsmadi MK, Omar KB, Noah SA, Almarashdah I (2009) Performance comparison of multi-layer perceptron (Back Propagation, Delta Rule and Perceptron) algorithms in neural networks. 2009 IEEE International Advance Computing Conference, IACC 2009, 7: pp 296–299Google Scholar
- Annadurai S, Shanmugalakshmi R (2011) Image transforms, fundamentals of digital image processing. Dorling Kindersley (India) Pvt. Ltd., pp 31–66Google Scholar
- Banerjee M, Kundu MK, Maji P (2009) Content- based image retrieval using visually significant point features. Fuzzy Sets Syst 160:3323–3341View ArticleGoogle Scholar
- Douze M, Jégou H, Singh H, Amsaleg L, Schmid C (2009) Evaluation of GIST descriptors for web-scale image search. In ACM International Conference on Image and Video Retrieval, pp 0–7Google Scholar
- Dubois SR, Glanz FH (1986) An autoregressive model approach to two-dimensional shape classification. IEEE Trans Pattern Anal Mach Intell 8(1):55–66Google Scholar
- Dunham MH (2009) Data Mining Introductory and Advanced Topics: Pearson Education, p 127Google Scholar
- ElAlami ME (2011) A novel image retrieval model based on the most relevant features. Knowl-Based Syst 24:23–32View ArticleGoogle Scholar
- ElAlami ME (2014) A new matching strategy for content based image retrieval system. Appl Soft Comput J 14:407–418View ArticleGoogle Scholar
- Flickner M, Sawhney H, Niblack W, Ashley J, Huang Q, Dom B, Gorkani M et al (1995) Query by image and video content: the QBIC system. Computer 28(9):23–32 IEEE View ArticleGoogle Scholar
- Gevers T, Smeulders AW (2000) PicToSeek: combining color and shape invariant features for image retrieval. IEEE Trans Image Proc Publ IEEE Signal Proc Soc 9(1):102–119View ArticleGoogle Scholar
- Hiremath PS, Pujari J (2007) Content based image retrieval based on color, texture and shape features using image and its complement. Int J Computer Sci Secur 1:25–35Google Scholar
- Irtaza A, Jaffar MA, Aleisa E, Choi TS (2013) Embedding neural networks for semantic association in content based image retrieval. Multimed Tool Appl 72(2):1911–1931Google Scholar
- Jalab HA (2011) Image retrieval system based on color layout descriptor and Gabor filters. 2011 IEEE Conference on Open Systems. pp 32–36Google Scholar
- Keerthi SS, Shevade SK, Bhattacharyya C, Murthy KRK (2001) Improvements to Plattʼs SMO Algorithm for SVM classifier design. Neural Comput 13:637–649View ArticleGoogle Scholar
- Kekre HB, Thepade S (2009) Improving the performance of image retrieval using partial coefficients of transformed image. Int J Inf Retr Ser Publ 2(1):72–79Google Scholar
- Kekre HB, Thepade S, Maloo A (2010) Image Retrieval using Fractional Coefficients of Transformed Image using DCT and Walsh Transform‖. Int J Eng Sci Technol (IJEST) 2(4):362–371Google Scholar
- Kekre HB, Thepade S, Das R, Ghosh S (2013) Multilevel block truncation coding with diverse colour spaces for image classification. In: IEEE-International conference on Advances in Technology and Engineering (ICATE), pp 1–7Google Scholar
- Kim WY, Kim YS (2000) Region-based shape descriptor using Zernike moments. Sig Process Image Commun 16:95–102View ArticleGoogle Scholar
- Li J, Wang JZ (2003) Automatic linguistic indexing of pictures by a statistical modeling approach. IEEE Trans Pattern Anal Mach Intell 25:1075–1088View ArticleGoogle Scholar
- Liu C (2013) A new finger vein feature extraction algorithm, In: IEEE 6th. International Congress on Image and Signal Processing (CISP), pp 395–399Google Scholar
- Luo H, Lina Y, Haoliang Y, Yuan YT (2013) Dimension reduction with randomized anisotropic transform for hyperspectral image classification. In: 2013 IEEE International Conference on Cybernetics, CYBCONF 2013, pp 156–161Google Scholar
- Madireddy RM, Gottumukkala PSV, Murthy PD, Chittipothula S (2014) A modified shape context method for shape based object retrieval. SpringerPlus 3:674. doi:10.1186/2193-1801-3-674 View ArticleGoogle Scholar
- Mehtre BM, Kankanhalli MS, Lee Wing Foon (1997) Shape measures for content based image retrieval: a comparison. Inf Process Manage 33:319–337View ArticleGoogle Scholar
- Mokhtarian F, Mackworth AK (1992) A theory of multiscale, curvature-based shape representation for planar curves. IEEE Trans Pattern Anal Mach Intell 14:789–805View ArticleGoogle Scholar
- Otsu N (1979) A threshold selection method from gray- level histogram IEEE transactions on systems. Man Cybern 9:62–66View ArticleGoogle Scholar
- Prakash O, Khare M, Srivastava RK, Khare A (2013) Multiclass image classification using multiscale biorthogonal wavelet transform, In: IEEE Second International Conference on Information Processing (ICIIP), pp 131–135Google Scholar
- Pratt W, Chen WH, Welch L (1974) Slant transform image coding. IEEE Transactions on Communications 22Google Scholar
- Rahimi M and Moghaddam ME (2013) A content based image retrieval system based on color ton distributed descriptors. Signal Image Video Process 9(3):691–704. http://dx.doi.org/10.1007/s11760-013-0506-6
- Ramírez-Ortegón MA and Rojas R (2010) Unsupervised evaluation methods based on local gray-intensity variances for binarization of historical documents. Proceedings—International Conference on Pattern Recognition, pp 2029–2032Google Scholar
- Raventós A, Quijada R, Torres L, Tarrés F (2015) Automatic summarization of soccer highlights using audio- visual descriptors. SpringerPlus 4:301. doi:10.1186/s40064-015-1065-9 View ArticleGoogle Scholar
- Shaikh SH, Maiti AK, Chaki N (2013) A new image binarization method using iterative partitioning. Mach Vis Appl 24(2):337–350View ArticleGoogle Scholar
- Shen GL and Wu XJ (2013) Content based image retrieval by combining color texture and CENTRIST, In: IEEE international workshop on signal processing, vol 1, pp 1–4Google Scholar
- Sridhar S (2011) Image features representation and description digital image processing. India Oxford University Press, New Delhi, pp 483–486Google Scholar
- Subrahmanyam M, Maheshwari RP, Balasubramanian R (2012) Expert system design using wavelet and color vocabulary trees for image retrieval. Expert Syst Appl 39:5104–5114View ArticleGoogle Scholar
- Subrahmanyam M, Wu QMJ, Maheshwari RP, Balasubramanian R (2013) Modified color motif co- occurrence matrix for image indexing and retrieval. Comput Electr Eng 39:762–774View ArticleGoogle Scholar
- Thepade S, Das R, Ghosh S (2013a) Advances in computing, communication and control. Image classification using advanced block truncation coding with ternary image maps, vol 361. Springer, Berlin, pp 500–509. doi:10.1007/978-3-642-36321-4_48 Google Scholar
- Thepade S, Das R, Ghosh S (2013b) Performance comparison of feature vector extraction techniques in RGB color space using block truncation coding or content based image classification with discrete classifiers. In: India Conference (INDICON), IEEE, pp 1–6. doi: 10.1109/INDCON.2013.6726053
- Thepade S, Das R, Ghosh S (2014a) A novel feature extraction technique using binarization of bit planes for content based image classification. J Eng. doi:10.1155/2014/439218
- Thepade S, Das R, Ghosh S (2014b) Feature extraction with ordered mean values for content based image classification. Adv Comput Eng 2014. doi:10.1155/2014/454876
- Valizadeh M, Armanfard N, Komeili M, Kabir E (2009) A novel hybrid algorithm for binarization of badly illuminated document images. 2009 14th International CSI Computer Conference, CSICC 2009, pp 121–126Google Scholar
- Walia E, Pal A (2014) Fusion framework for effective color image retrieval. J Vis Commun Image Represent 25(6):1335–1348Google Scholar
- Walia E, Vesal S, Pal A (2014) An Effective and Fast Hybrid Framework for Color Image Retrieval. Sens Imaging 15:93. doi: 10.1007/s11220-014-0093-9
- Wang X, Bian W, Tao D (2013) Grassmannian regularized structured multi-view embedding for image classification. IEEE Trans Image Process 22(7):2646–2660View ArticleGoogle Scholar
- Yanli Y and Zhenxing Z (2012) A novel local threshold binarization method for QR image, In: IET International Conference on Automatic Control and Artificial Intelligence (ACAI), pp 224–227Google Scholar
- Yıldız OT, Aslan O, Alpaydın E (2011) Multivariate statistical tests for comparing classi-fication algorithms. Lect Notes Comp Sci, vol 6683, Springer, Berlin, pp 1–15Google Scholar
- Yue J, Li Z, Liu L, Fu Z (2011) Content-based image retrieval using color and texture fused features. Math Comput Model 54:1121–1127View ArticleGoogle Scholar
- Zhang D, Lu G (2003) A comparative study of curvature scale space and Fourier descriptors for shape- based image retrieval. J Vis Commun Image Represent 14:39–57View ArticleGoogle Scholar
- Zhang D, Lu G (2004) Review of shape representation and description techniques. Pattern Recogn 37:1–19View ArticleGoogle Scholar
- Zhao C, Li X, Cang Y (2015) Bisecting k-means clustering based face recognition using block-based bag of words model. Optik Int J Light Electron Optics 126(19):1761–1766Google Scholar
- Zhu Q, Shyu M-L (2015) sparse linear integration of content and context modalities for semantic concept retrieval. IEEE Trans Emerg Top Comput 3(2):152–160View ArticleGoogle Scholar