Analyzing deep features for trademark retrieval

The rapid rise in the amount of trademark applications and trademark infringements has led the trademark retrieval (TR) to become an important and formidable task to solve. Existing studies based on hand-crafted features show unsatisfying performance. Taking the popularization and increasing success of the deep learning methods into consideration, in this work, many well-known Convolutional Neural Network (CNN) models are applied to the TR problem. Models are tested using a large scale trademark dataset in contrast with the previously proposed solutions, and their failure points are discussed in this study. For the problems that can be encountered, solutions such as fine-tuning, distance metric learning, using CNN features locally, and making them invariant to aspect ratio of the trademark are suggested.