Pavol Dubovec
Supervisor(s): prof. Ing. Adam Herout Ph.D.
Brno University of Technology
Abstract: This paper presents a new technique for locating a photograph within a larger one, with the aim of enhancing the speed and accuracy of conventional methods. The proposed technique utilises a CNN architecture to extract multiple embeddings from the query image, which are then used to perform an approximate search within a database of embeddings from the large photograph. Two main models were trained on a large dataset. The first model used a triplet loss function, while the second model used a cross-entropy loss function. Conventional methods were used to determine the location of the images in the training set and to generate a large image. A database of embeddings was created by partitioning the large photograph with a certain sampling frequency (in pixels) using the trained model. The database is queried for K-nearest sub-query embeddings. These embeddings are generated by partitioning the query image into equal-sized pieces as CNN inputs. The optimal homography model is determined through random sampling based on the positions of four sub-query images and their corresponding positions in the large image. The model homography with the lowest harmonic mean embedding distance is selected as the resulting position. The method demonstrates satisfactory accuracy and good speed on the generated test datasets. The best model achieved a top-1 accuracy of 97.71% and a top-3 accuracy of 99.17%. Future research will investigate the method's performance with increasing surface heterogeneity, the potential for automating video retrieval to obtain a large dataset of photos, and its effectiveness for photo localization in cases where conventional methods fail due to a lack of key points.
Keywords: Computer Vision, Image Processing
Full text: Year: 2024
