Development of robust feature matching techniques in images using deep learning

Computer vision techniques widely utilize feature points detected in images. Using these feature points, useful consistencies can be detected between pairs of images. By leveraging these similarities, successful results can be achieved for various applications such as image matching, object recognition, image stitching, creating image mosaics, and object tracking. When matching feature points detected in pairs of images, the distances between the points in the feature space are used as the basis. If the proximities of the features closest to each other in the feature space are sufficiently distinctive, these features are considered putative matches. However, these putative matches often cannot entirely exclude incorrect matches. To address this, iterative algorithms are used in the literature to achieve geometric consistency that includes the maximum number of putative matches. Thanks to the geometric consistency obtained, incorrect matches in the putative matches are eliminated. The main issue with this method is the practical impossibility of finding a repetition number that guarantees success for all image pairs. Following the effective results achieved by deep learning methods over alternatives in many computer vision problems in the literature, solutions that use artificial neural networks trained using deep learning have become prevalent for the feature matching problem, as they have in most computer vision and image processing problems. In this thesis, solutions to the problem of relative camera pose estimation have been developed using deep learning networks that perform stable feature matching. First and foremost, the n-to-n framework, which forms the basis of all studies in the literature, was examined. This n-to-n framework was observed to operate on a set-type input consisting of the coordinates of features in putative matches. The challenges faced by studies in the literature using this framework in deriving global and local contexts from putative feature matches were analyzed. As an alternative to the n-to-n framework, a one-to-one framework was proposed. In the n-to-n framework, each batch contains data belonging to a single image pair, whereas, in the proposed one-to-one framework, each batch also includes data belonging to only a single image pair. The one-to-one framework allows the use of a dedicated context channel for each putative match of the image pair. Moreover, it is possible to specifically sort the input rows in this context channel for each putative match. Within the scope of the thesis, a variety of fundamental artificial neural networks with different architectures using the one-to-one framework were generated. The one-to-one artificial neural networks developed, along with the successful n-to-n ones in the literature, were trained on Tensor Processing Units (TPUs). Multiple loss functions were utilized during the training. The minimum average precision was used as the performance metric. According to the results, the artificial neural networks designed for the one-to-one framework outperformed all but one of the n-to-n neural networks from the literature by up to 30\%. Additionally, before working on the main topic of the thesis, a case study was conducted to increase familiarity with deep neural networks. For this purpose, artificial neural networks were developed to estimate age categories using portrait photographs. The portraits in the dataset were divided into six classes. It was observed that a 6-layer artificial neural network from the literature performed better than the 18 and 34-layer residual neural networks. It was concluded that the 18-layer and 34-layer residual networks might have been unnecessarily deep for the 6-class age estimation problem, leading to overfitting, or that the dataset lacked sufficient diversity to clearly demonstrate the advantages of deeper networks.