Deduplication of Advertisement Assets Using Deep Learning Ensembles

Creative image deduplication is the process of detecting similar images among digital advertisement assets. We evaluated a novel ensemble of DenseNet based Siamese Network and Scale Invariant Feature Transformation (SIFT) which is able to outperform both individually. Our approach also outperforms a state-of-the-art model on the desired metrics in the adtech world. The evaluation was carried upon the advertisement assets of a major ad-network. Due to the difficulty in acquiring labelled images for the training dataset, we also present a novel strategy for semi-automatically generating labels.