Component-Based Attention for Large-Scale Trademark Retrieval

Abstract

The demand for large-scale trademark retrieval (TR) systems has significantly increased to combat the rise in international trademark infringement. Unfortunately, the ranking accuracy of current approaches using either handcrafted or pre-trained deep convolution neural network (DCNN) features is inadequate for large-scale deployments. We show in this paper that the ranking accuracy of TR systems can be significantly improved by incorporating hard and soft attention mechanisms, which direct attention to critical information such as figurative elements and reduce attention given to distracting and uninformative elements such as text and background. Our proposed approach achieves state-of-the-art results on a challenging large-scale trademark dataset.

MTRNet: A Generic Scene Text Eraser

“…Text removal methods are categorized into one-step [5] and two-step [8] based approaches. Here, related works on these two types of approach are discussed.…”

“…Early two-step approaches [4,9] are mainly based-on primitive hand-crafted text-detection and inpainting algorithms. A recent two-step approach with more accurate text localization is proposed by Tursun et al [8]. They proposed pixel level text segmentation for text localization, however, inpainting is implemented by replacing text pixels with the most-frequently occurring neighborhood background color around the text region.…”

“…The task has drawn the attention of the computer vision community due to its potential value for privacy protection [5,13], as text removal approaches automatically remove private information such as license plate numbers, addresses, and names from images shared via social media. It is also helpful for other applications including image editing, image restoration, [5,11] and image retrieval [8].…”

Unsupervised Trademark Retrieval Method Based on Attention Mechanism

“…At present, most trademark retrieval methods are based on deep learning extract trademark features by a supervised way [ 7 ]. Perez et al proposed a retrieval of trademarks through the combined VGG network [ 1 ] with supervised training, Tursun et al [ 9 ] removed the text of trademarks and combined soft and hard attention mechanisms to direct attention to key information. Lan et al [ 10 ] proposed a method to extract uniform Local Binary Pattern (LBP) features from the feature map of each convolutional layer feature, and achieved good results in both METU and NPU trademark datasets.…”

Learning Regional Attention Over Multi-Resolution Deep Convolutional Features For Trademark Retrieval

“…Later, LSTR using content-based image retrieval (CBIR) algorithms have been used thanks to it's efficiency and accuracy. Hand-crafted features based-on shape, color or texture were developed for early CBIR-LSTR systems [2,3]. With the rise of deep learning, off-the-shelf deep features have been applied for LSTR, demonstrating higher accuracy and efficiency compared to traditional hand-crafted features.…”