An Integrated approach to CBIR using multiple features and HSV Histogram

— An efficient search for semantically relevant images has always been thirst in computer vision and processing specially in large scale image retrieval. We propose an integrated approach for fast and effective image retrieval system using multiple features and hsv histogram. Relevance feedback allows user interaction to improve the performance. Features used in this work are improved lbp and modified fourier descriptors, plays vital role in effective retrieval. Experimental results on CALTECH-101 and MPEG CE shape 1 datasets proves that our framework provides better retrieval efficiency compared to the state of art methods.

. Block diagram of the Proposed Image retrieval system A. Shape features The shape of the objects in images gives important information in retrieval. Fourier descriptors are one of such shape feature which is fast and effective. Wide variety of shape descriptors exist, classified into contour based and region based descriptors. Contour based descriptors give the boundary information in detail but they fails in case of disjoint shapes. Region based descriptors use the pixel information with in the entire region. Several descriptors such as hu moments, zernike moments and legendre moments were discussed. Shape variations are represented by the sample points along the three most dominant principle axes in the feature space [9]. Zernike moments gives efficient image retrieval, but it has two disadvantages, the first one is it has redundant features at each repetition order. The second one is it cannot compute features in radial directions. Fourier descriptors are powerful features for the recognition of two-dimensional connected shapes as well as for broken shapes [10]. So we propose a Modified Fourier Descriptor (MFD), which provides multi resolution analysis. In general, FT (Fourier Transform) is used in shape analysis. But the features which are calculated from FT are not rotation invariant. In order to achieve the rotation invariance the polar form is used. For the given image f(x,y) the modified polar transform is given by eq. (1). (1) Where ρ and ϕ are radial and angular frequencies and , , where , and . (2) Where area is the area of the boundary with in the shape resides, m is the maximum no of radial frequencies and n is the no of angular frequencies selected. For efficient implementation few MFD features can be used. In our framework we used 36 features. The similarity comparison between the features is done by using ecludian distance. Fig. 2(a) shows the query image and It is clearly observed that the FFT is different for the two images, but there is no difference in the polar forms. So the rotational invariance is achieved by using the proposed modified fourier descriptor.

B. Texture Feature extraction
The texture is a powerful low-level feature for image search and retrieval applications [1]. In the field of computer vision, Local binary patterns are one of the best texture based visual descriptor. LBP is used to analyse the texture of images in the dataset and retrieve the same more appropriately. The computation of an LBP can be described as; the labels for each image pixels are formed by thresholding the surrounding 3 * 3 pixels with the centre value in order to generate the binary number to each pixel of an image. Finally the histogram plotted for all 256 pixels is obtained for texture description and analysis. The efficiency of the original LBP operator is enhanced by two main extensions namely discriminative capability and its robustness. The local structures of an image are defined by LBP.  5 shows the next modification, ILBP is improved by operating on neighbourhoods of different radius from the circle. The notation in Fig.5 represents (P,R) where P denotes the no. of neighbourhood points on a circle of radius R. The ILBP (Improved LBP) operator makes the comparison among all the pixels with their mean intensity values. This is given in Fig.3. This above mention calculation was performed only on 3x3 matrixes. To improve the performance and making rotational invariant the Modified LBP operator is proposed in this paper. In Fig. 6, ILBP is represented as the modified version of original LBP operator. It performs the comparison among the neighbourhood pixels and central pixel along with the calculation of grey-value differences. Fig. 7 represents the modified LBP (MLBP) operator. In this the feature extracted consists of several layers of LBP codes along with the grey-value differences between the neighbouring and central pixels. Firstly all the grey-scale difference values are encoded to its binary values representation. Later all the binary values formed in a layer expresses as modified LBP. Hence, compared to first LBP layer the information encoded in additional layers tends to be more discriminative. This distinguishes the robustness of the operator. MLBP highlights the dimensionality of the feature to a greater extent.
For a given pixel(x,y) in the given input image the LBP is calculated by eq.(4). where gp is the gray level intensity of the particular pixel and gc is the gray level value with in the 3*3 neighbourhood of the centre pixel. The rotational invariance can be achieved by eq. (5).
Where ROR performs bit wise circular right shift in the specified neighbourhood. The features from the rotation invariant lbp are stored in the feature vector. C. HSV histogram Hsv histogram is the feature which is used for the color image database. The rgb images from the database are converted into hsv space by using the following expressions.  ( 1 0 ) The quantized histogram values of the query image are compared with database images by using chi-square distance. All the three features are stored individually and the similarity matching distance is calculated. The similarity matching for the shape and texture features is done with euclidean distance. The final distance with which the images are sorted is given by D final = D mfd + D lbp + D hsv . With this distance the images are indexed in ascending order and displayed.

III. SIMILARITY MATCHING
The similarity matching between the features of query image and the database images is calculated by using the following distance metrics. Chi-square distance: The chi-square distance between the feature vectors of query image (Qi) and database images (Dbi) is given by eq. (11).

IV. RESULTS
The proposed framework is tested with two different databases CALTECH-101 and MPEG CE-shape 1. MPEG-7 CE shape 1 dataset consists of 70 different categories of images each with 20 similar images. Fig. 8. shows the top ten retrieval results for different classes of images. The experiments were conducted by taking 20, 30 and 50 different classes randomly from the dataset. The precision is calculated by using the following equation.   Fig. 9. Fig.10 shows the average efficiency of the proposed framework for different classes of CALTECH dataset with different number of images. The maximum retrieval efficiency for CALTECH-101 dataset is 84% and for the MPEG-7 dataset is 92%.   Table I shows the average retrieval efficiency for MPEG-7 CE shape I dataset. In this different categories of images were considered and the proposed algorithm was tested. The maximum retrieval performance is 92%, which is optimum compared to the existing methods.
V. CONCLUSION In this paper, an integrated and effective image retrieval system is proposed with hsv histogram, modified fourier descriptors and improved local binary patterns. The modified fourier descriptors are used in polar form in order to achieve the rotation and translation invariance. The shape descriptors are extracted from the modified fourier descriptors. The texture features are extracted from improved LBP. The hsv histogram is used for color feature matching. The proposed approach is an integrated approach as it can be used for both rgb images and gray-scale images. The retrieval efficiency is optimum for both the test databases considered. The efficiency is 88% and 72% which is greater than the existing methods.