Multimodal Image Retrieval using PLSA and Microstructure Descriptor

PLSA (Probabilistic Latent Semantic Analysis) and SIFT (Scale Invariant Feature Transform) are widely used techniques that have been known as state of the art of multimodal image retrieval. However, for a gray-scale image, SIFT produces a big number of keypoints, where each keypoint has a 128 dimensions feature vector. SIFT does not store any information about the image color. This leads to an enormous amount of descriptors especially when it is applied in a big database like Flickr. On the other hand, Micro Structure Descriptor (MSD) represents a full color image as a 72 dimensions feature vector. Furthermore, MSD comprises the information about colors, textures and shapes. This paper presents a PLSA based multimodal image retrieval system using MSD feature extraction algorithm. In the evaluation we compare our proposed system to PLSA based multimodal image retrieval system using SIFT feature extraction algorithm. The extensive experiment results show that PLSA-MSD image retrieval system is more efficient than PLSA-SIFT, accounted for 300% faster in terms of computational speed. The results imply that PLSA-MSD is suitable for big databases.