Representation based Few-Shot Learning for Brand-logo Detection

Abstract

In this paper, we propose an attention-net based few-shot object detection (AN-FSOD) model for brand-logo detection and recognition. With the fact that brand-logo detection has many distinct properties: tiny objects, similar brands, and adversarial images, most of the current FSOD approaches, motivated by meta-learning, metric-learning and transfer learning techniques, typically perform less-effective due to the difficulties on target region allocation. The proposed AN-FSOD aims to locate the region of the brand-logo targets, achieved by a well- trained attention-net, therefore providing an explicit feature maps for detection and classification. An end-to-end feature extractor and target detector model is established, implementing with a simultaneous parameter fine-tuning with respect to the few-shot dataset. Extensive simulations have confirmed that the proposed AN-FSOD gains significantly better performance than the vanilla FSOD model and the majority of the feature extractor aligned model on a public brand-logo dataset.