Research On Logo Detection Technology Integrating Multi-Level Features

With the high quality development of the economy,more and more companies are emerging.Logo is an integral part of a company as a presentation of its corporate image.Companies have invested significant resources in creating unique Logos.The diversity and uniqueness of Logos has made Logo detection technology increasingly important in real life applications and has a wide range of applications.On the one hand,Logo detection is able to detect trademark infringement timely and protect the legal rights of the company.Its reasonable regulation of trademark infringement and favourable protection of intellectual property rights can strongly build a healthy and green market order.On the other hand,Logo detection enables the accurate identification of Logos and the rapid access to product information,helping consumers to obtain a fuller understanding of product content.As a result,Logo detection has profound significance for our real-world applications.Logo detection,as a subdivision of object detection,has the key task of finding the location and category of Logos in an image.In practical application scenarios,traditional detection methods are difficult to meet realistic needs due to problems in Logo detection such as multi-scale,small target,image quality,distortion and data imbalance.As a result,the construction of deep learning Logo detection methods that can meet practical needs has become a hot topic in the field.By identifying the properties of Logo images and the performance limitations of detection algorithms in Logo-related works,this thesis improves the baseline model to reach higher detection capability.The work in this thesis includes the following:For some Logo images with large aspect ratio characteristics and multi-scale detection problems,a novel Logo detection method is proposed in this thesis based on neural structure search and long-range dependence involution.First,by rethinking the intrinsic principles of convolution,a long-range involution module is proposed and applied to the region proposal network.It alleviates the detection difficulties arising from large aspect ratios by a new operator and a self-attentive mechanism.Then,a multi-level representational neural structure search is proposed and applied to multi-scale Logo detection,which introduces six information paths to build a richer topology,achieves the fusion of semantic information and location representation,and enhances the cross-layer interaction between features.In addition,the modelling capability is enhanced by using an adaptive Ro I Pooling module to achieve adaptive feature learning for objects of different shapes.In order to solve the problem of detection difficulties caused by few extractable features and Logo aggregation of small Logos,a new Logo detection network based on the transmission of cross-direction features and task decoupling is proposed in this thesis.By constructing a cross-direction feature pyramid using horizontal and vertical transmission,the feature information is enhanced to achieve more effective feature fusion within a small Logo region.The horizontal transmission uses iterative feature pyramids,followed by vertical transmission focusing on extracting balanced semantic representations.To solve the aggregation problem of small Logos,the network introduces a multi-frequency task decoupling detection head.To address the difficulties of multi-objective regression,this thesis proposes a multi-frequency attention convolution branch,which combines the advantages of discrete cosine transform and convolution.The fully connected layer is used to construct fully connected classification branches as it has a strong ability to distinguish between Logos categories.In summary,the Logo detection algorithms integrating multi-level features are proposed in this thesis,and the proposed detection networks are experimental evaluated on several Logo datasets of different sizes.In comparison with several existing deep learning-based detection methods,the experiment results confirm the validity of the proposed networks in this thesis.