The effectiveness of LSI-based CBIR with image noise using wavelet-based texture

Content-based image retrieval (CBIR) technique retrieves relevant images based on extracted features from image contents. Latent semantic indexing (LSI) is used as a semantic model in the CBIR field. This paper investigates the capability of LSI-based CBIR in dealing with different types of image noise, and the impact of noise on the retrieval results. To construct the feature-image matrix (FIM) in the proposed LSI framework, three wavelet-based methods are used to extract texture feature: Gabor wavelet, Daubechies wavelet, and wavelet moments. The performance of the proposed system is evaluated by a predefined accuracy measure. The results show that the LSI-based CBIR achieves a high level of accuracy with the original image database, and still performs very well in dealing with different types of image noise.