Visual Search using Dimensionality Reduction and Improved Similarity Models

In times of growing databases of digital images, visual search is becoming increasingly important. Coping with large amounts of unannotated visual data requires methods to search and sort automatically. Visual search using global image representations is one possible solution. This thesis introduces approaches for a better local feature representation and more efficient search by decorrelation and dimensionality reduction, an improved spatial image partitioning for image retrieval of multi-object images and better similarity metrics for global image representations.
Orthogonal transforms applied to the local feature descriptors show improved performance in retrieval due to the decorrelation and better representation of the descriptors' inherent information. Additionally, the dimension reduction helps in saving computation time and memory which is especially helpful for large-scale applications.
Spatial partitioning deals with the loss of spatial information when a global image representation is generated. Using visual saliency, objects can be detected and the image can be spatially partitioned using the object locations.
Also, the concept of metric learning is introduced for improved similarity measurements. Weighted Approximate Rank Component Analysis was applied to learn a better similarity metric. WARCA also has the property of introducing a dimensionality reduction by transforming the data into a lower dimensional space for the distance computation.
The solutions presented in this thesis are a step towards a better performance for global image representations used in image retrieval application. An improved data representation and trained similarity metrics help to increase the performance while lowering the complexity and memory consumption of these applications.