Efficient colour texture image retrieval by combination of colour and texture features in wavelet domain

C. Bai, W. Zou, K. Kpalma, J. Ronsin
2012 Electronics Letters  
Efficient colour texture image retrieval by combination of colour and texture features in wavelet domain. A new approach for image retrieval by combination of color and texture features is proposed. This approach uses the histogram of feature vectors which are constructed from the coefficients of some subbands of wavelet transform and chosen according to their intrinsic characters. K-means algorithm is used to quantize feature vectors. The experimental results both on small size database (40
more » ... sses of textures) and large size database (167 classes of textures) show that, compared with the state-of-the-art approaches, the proposed approach can get better retrieval performance. Introduction: Content-based image retrieval (CBIR) is an active research field in pattern recognition and computer vision. Color and texture are two important features that are used in CBIR. Using the combination of both features provides better performance than that of color or texture alone. For example, in [1], red, green and blue (RGB) channels of color images are treated as three respective pseudo gray-level images and Gabor filters are applied on these three images to extract features. In our previous work [5], texture features are constructed from the AC coefficients of Discrete Cosine Transform (DCT) and color features are constructed from the DC coefficients. As wavelet is widely used as an efficient tool for extracting features, some researchers have presented image retrieval methods based on wavelet in recent years. So in [2], color features are represented by 2D histogram of CIE Lab chromaticity coordinates and texture features are extracted by using Discrete Wavelet Frames (DWF) analysis. In [3], RGB images were firstly transformed into HSV model. The color feature is represented by the autocorrelogram of wavelet coefficients extracted from Hue and Saturation components, and the first and second moments of the BDIP (block difference of inverse probabilities) and BVLC (block variation of local correlation coefficients) for each subband of Value component is used as texture feature. A recent approach is presented in [4], in which, the wavelet coefficients in RGB color channels are modeled by multivariate Laplace distribution and Student-t distribution. Other than mentioned methods, this paper presents a new method for color texture image retrieval combining color and texture features in wavelet domain. This method constructs the color and texture features from the coefficients of some subbands of wavelet transform.
doi:10.1049/el.2012.2656 fatcat:5pxqft5lkjhmbd2jsoqfxyrdfi