YDA imgelerde görsel benzerlik ve ton eşlemeye uygulamaları

Assessing visual similarity between images is important for many computer vision applications. So far, investigations on visual similarity have been confined to low dynamic range images. However, recently, there is a growing interest to high dynamic range (HDR) imaging. In this thesis, the aim is to shed light on visual image similarity for HDR images by following an experimental approach. To this end, a user experiment is conducted through a novel web-based interface, in which the participants assess the pairwise similarity of HDR images. The data collected through this experiment is used to evaluate a set of image features with respect to their correlations with the participants responses. A combined feature as a linear combination of individual features is defined, and its coefficients are learned using metric learning. The learned combined feature as compared to individual features correlated better with the participants' responses. Among individual features, deep learning features are found to correlate better than others, supporting that higher level features are better than lower level ones. Using the learned similarity, i.e. combined feature, a style-based tone mapping algorithm is proposed that successfully imparts a user-defined style to various HDR images determined to be similar with respect to the proposed metric.