Fast Depiction Invariant Visual Similarity for Content Based Image Retrieval Based on Data-driven Visual Similarity using Linear Discriminant Analysis
Yaya Wihardi, Wawan Lulus Setiawan, Eki Nugraha · IOP Conference Series Materials Science and Engineering · 2018
On this research we try to build CBIRS based on Learning Distance/Similarity Function using Linear Discriminant Analysis (LDA) and Histogram of Oriented Gradient (HoG) feature. Our method is invariant to depiction of image, such as similarity of image to image, sketch to image, and painting to image. LDA can decrease execution time compared to state of the art method, but it still needs an improvement in term of accuracy. Inaccuracy in our experiment happen because we did not perform sliding windows search and because of low number of negative samples as natural-world images.