Logo Detection Methods Based On Deep Learning

In recent years,with the rapid development of technology and deepening globalization,intellectual property has become a key factor in innovation and competition.As a result,protecting intellectual property is becoming increasingly prominent.As a critically important form of intellectual property,logos not only act as the brand symbol but also can enhance its competitiveness in the fierce market competition.Therefore,logo protection becomes particularly important.Intelligent logo detection,which predicts the category and location of logos in images,is an important component in infringement detection systems for protecting intellectual property.However,there are a huge number of brand logos over the world,many of which are difficult to distinguish because of similar appearance.In addition,many logos are too small and lack discriminative features,making them more likely to be missed.To address these issues in logo detection,this thesis proposes a series of solutions based on deep learning techniques and verifies the effectiveness of the methods through detailed experiments.The main work in this thesis includes:(1)Cascaded logo detection for class disambiguation.Existing detection algorithms usually have difficulty in distinguishing similar logos.This thesis proposes a two-stage detection framework to eliminate classification ambiguity.The framework consists of cascaded logo localization and class verification stages.A lightweight detector is designed to localize potential logos in the localization stage.Then,in the verification stage,a verification network is designed to improve the logo classification accuracy within overlapping proposal groups by exploring more discriminative features.The proposed method has been validated on two logo detection datasets,showing its effectiveness in distinguishing similar logos and removing overlapping false positive predictions caused by classification ambiguity.(2)Sample-balanced training strategy for small logo detection.To address the problem of low detection accuracy and high missing rate of small logos,this thesis proposes a balancing training strategy for multi-scale logo samples.The proposed training strategy first adaptively stitches training images according to the training loss while preserving the size of small logos during stitching to balance the training loss contribution of logos of various scales.Secondly,it adopts adaptive label assignment methods for logos of different scales to ensure sufficient positive samples for logos of each scale.The proposed method has been validated on two publicly available logo detection datasets and shows its effectivness in improving the detection accuracy and reducing the rate of missed small logos.Overall,this thesis conducts research on logo detection and proposes a number of effective solutions.The proposed cascaded logo localization and class verification stages can eliminate the classification ambiguity,while the combination of the training strategy based on self-feedback and multi-scale label assignment method greatly improves the detection accuracy of small logos.These proposed methods provide valuable references for relevant research in logo detection and can help protect intellectual property.