Content-based image retrieval through extended normalised ridgelet-fourier and multi-resolution joint auto correlograms

Three methods for Content-based Image Retrieval (CBIR) have been proposed which are the Multi-resolution Joint Auto Correlograms (MJAC) as a colour-based approach, the Extended Normalised Ridgelet-Fourier (ENRF) as a shape-based approach, and the Integrated Colour-Shape descriptor (ICS) which combines the proposed colour and shape approaches. Colour Auto Correlogram (CAC) is one of the most promising colour-spatial descriptors. However, the conventional CAC and most of its advancements are sensitive to scale, are only based on a single feature, and are computed in the spatial domain. The MJAC is introduced as an extension of several CAC advancements by allowing multiple image features to represent an image rather than just colour and extracting them at different image sub-bands to provide different physical structures of the image in the frequency domain. The Ridgelet transform (RT) is performed on the RGB colour space and the grey-scale version of the images to provide the multiresolution levels. The colour feature is extracted from the Ridgelet coefficients of the RGB colour space while other image features such as gradient magnitude, rank,and texturedness are extracted from the Ridgelet coefficients of the grey-scale image. Each of these image features is quantised and an auto correlogram is then performed on the respective quantised image feature coefficients. Retrieval experiments conducted on the 1000 SIMPLIcity image dataset have shown that the proposed MJAC is able to obtain a higher precision rate of 78.52% compared to the benchmark methods. Complicated shapes can be effectively characterised by using a description with multiple resolutions. One popular multi-resolution method is the RT which has enjoyed very little exposure in describing shapes for CBIR. Apart from that, many of the existing RTs are only applied on images of sizeM × M . For M × N sized images, they need to be made M × M or segmented into M × M sub-images prior to processing. Furthermore, a different set of rho and theta for the Radon transform parameters need to be utilised according to the image size which results in computational complexity. Therefore, the ENRF has been proposed to tackle the above-mentioned issues regarding the previous work on Ridgelet descriptors by introducing a shape descriptor based on RT which is able to handle images of various sizes. The shape descriptor can then be applied in content-based shape retrieval applications. The utilisation of the ellipse template for better image coverage and the normalisation of the RT are introduced. For better retrieval, a template-option scheme is also introduced. The proposed ENRF has been tested on 1400 standard MPEG-7 CE-1 B image dataset and the retrieval effectiveness obtained by the proposed method is higher than the comparable methods, which is equivalent to 55.02%. The performance of the proposed colour and shape features can be further improved through feature fusion. ICS has been introduced by providing a scheme to integrate the proposed colour and shape descriptors to boost the performance of CBIR. Experimental results on 100 SIMPLIcity image dataset have shown that the proposed combination scheme is able to achieve a precision rate of 53.50%, slightly higher than the proposed colour-only and shape-only descriptors.

Preview PDF FSKTM 2011 35R.pdf Download (553kB) | Preview

Actions (login required)

View Item