Invariant Image Retrieval Using Wavelet Maxima Moment

There is a high demand for eeective and precise tools for users to search, browse, and interact with image databases and do so in a timely manner. Automatic feature extraction is a crucial part for any such retrieval systems. Current methods for feature extraction suuer from two main problems: rst, many methods do not retain any spatial information, and second, the problem of invariance with respect to standard transformation is still an unsolved problem. On the other hand, wavelets has been shown to be a powerful and eecient mathematical tool to process visual information at multiple scales. Some recent image retrieval systems use spatial information and visual features represented by dominant wavelet coeecients 1]. In addition, the underlying multiresolution mechanism of the wavelet decomposition allows the retrieval process to be done progressively. In this work, to cure the plague problem of translation variance with wavelet basis transform while keeping a compact representation, the wavelet transform modulus maxima 2] is employed. Essentially, wavelet maxima is obtained via an adaptive sampling scheme of the continuous wavelet transform along the time axis. Therefore the sampling grid is automatically translated when the signal is translated. Wavelet maxima has been shown to be very eeective in characterization of images from multiscale edges. Therefore feature extraction based on the wavelet maxima transform captures well the edge-based and spatial layout information which are likely the key features on an image query. The reconstructed power of wavelet maxima indicates the signiicancy of this representation. In addition, the "denoising" facility in the wavelet maxima domain is exploited to achieve the robustness in retrieving images which contain interested objects against various image backgrounds. To measure the similarity between wavelet maxima representations, which is required on the context of image retrieval systems, the diierence of moments is used. Normalized central moments are eeciently computed for each scale of the wavelet maxima transform. As a result, each image is indexed by multiscale vectors in feature spaces. Moreover, those moments are invariant with respect to both translation and scaling. Evaluation of the proposed method was performed on both synthetic and natural image collections. The rst test collection is composed of 15 images of basis shapes at diierent locations and scales, contaminated by Gaussian noise. The second test collection consists of 1000 color images from the Corel Professional Art Collection including landscapes, building, people, drawing , etc. The preliminary result shows that the proposed …