Content Based Image Retrieval (CBIR) for Brand Logos

This thesis explores the problem of automatically detecting the presence of logos in general images. Brand logos carry the goodwill of a company and are considered to be of high value in the corporate world, and thus automatically determining whether or not a logo is present in an image can be of interest for companies that wish to protect their brand. The problem of automated logo detection is inherently complex, but is further complicated through intentional obfuscation of logo images, for example by color shifting or other slight image modifications that leave the logo intact and easily recognizable by a human, but difficult to determine through automated techniques. For our use case, we are interested in leveraging the basic Content-based image retrieval (CBIR) approach to determine whether or not a brand logo is present within a larger image. CBIR systems retrieve images based on their similarity to a given query image/images. For example, this enables a user to retrieve images that contain dogs by entering an image with a dog in it. Of the multitude works done in this field, most use supervised and only a few use self-supervised machine learning techniques. For our approach presented in this thesis, we utilize an Autoencoder, which is one of the self-supervised techniques. It is able to generate compressed representations of the original image and its variant exposed to simple transformations such as Gaussian noise and near to zero degree rotations. It, however, is not able to do so for complex transformations/noises. This thesis addresses this problem by introducing a novel autoencoder based architecture that takes the encoding technique of an autoencoder one step further and produces a semantic rich compressed latent representation for the transformed logo that is similar to its original logo. This research work has achieved significant improvement over the general autoencoder architecture, and over current CBIR-based approaches for logo recognition