An unsupervised genetic algorithm framework for rank selection and fusion on image retrieval

Abstract

Despite the major advances on feature development for low and mid-level representations, a single visual feature is often insufficient to achieve effective retrieval results in different scenarios. Since diverse visual properties provide distinct and often complementary information for a same query, the combination of different features, including handcrafted and learned features, has been establishing as a relevant trend in image retrieval. An intrinsic difficulty task consists in selecting and combining features that provide a higheffective result, which is often supported by supervised learning methods. However, in the absence of labeled data, selecting and fusing features in a completely unsupervised fashion becomes an essential, although very challenging task. The proposed genetic algorithm employs effectiveness estimation measures as fitness functions, making the evolutionary process fully unsupervised. Our approach was evaluated considering 3 public datasets and 35 different descriptors achieving relative gains up to +53.96% in scenarios with more than 8 billion possible combinations of rankers. The framework was also compared to different baselines, including state-of-the-art methods.