Oran testi ile karşılıklı en yakın komşu aramasına dayanan öğrenmeye dayalı yeni bir görüntü eşleştirme yaklaşımı

This thesis proposes a novel image matching method that utilizes learned features extracted by an off-the-shelf deep neural network to obtain a promising performance. The proposed method simply uses a pre-trained VGG architecture as a feature extractor and does not require any additional training to improve matching. Inspired by well-established concepts in the psychology area, such as the Mental Rotation paradigm, an initial warping step is also performed by the help of a preliminary geometric transformation estimate. The matching estimates are based on dense matching using Mutual Nearest Neighbor Search with Bidirectional Ratio Test (MNNSwBRT) at the terminal layer of VGG network outputs of the images. After this initial alignment, the same approach is repeated again at every network level between reference and aligned images in a hierarchical manner to reach a good localization and matching performance. By comprehensive experiments, five classical and four learning-based methods in the literature are also compared while optimizing a single parameter, and it is shown that the proposed method achieves the state-of-the-art performance. As a result of a fair comparison, the experimental results on HPatches dataset reveal that the performance gap between classical and learning-based methods is not that significant as reported in most of the previous studies. Hence, one can conclude that our proposed method, which uses only a pre-trained network and ratio test, outperforms most well-trained learning-based methods.