Research On Trademark Retrieval Solution Based On Instance Discrimination

Intellectual property as a concentrated expression of innovation,its protection has risen to a national strategy.Trademark protection,as an important component of intellectual property protection,escorts corporate brand value.In recent years,intelligent trademark retrieval has become an important auxiliary means for collecting facts in the law enforcement process by virtue of its convenient and rapid advantages.However,due to its feature learning method and limitations in data collection,the current trademark intelligent retrieval system often has poor retrieval results and difficult data labeling,which seriously hinders its application in trademark enforcement scenarios.Facing the practical problems in trademark retrieval,this paper combines unsupervised feature learning and image generation technology to propose two methods to solve the practical problems of trademark retrieval system,and provide an effective auxiliary tool for actual trademark enforcement.Firstly,aiming at the performance and efficiency improvement of unsupervised trademark feature learning,this paper explores the sampling strategy in noise-contrastive estimation of instance discrimination algorithm,and proposes a sampling method based on relative neighbour sampling for noise-contrastive estimation to extract better trademark feature.Then a method of generating trademark images applied to trademark images is proposed to compensate for the lack of richness in the data sampling process.Then,a retrieval method based on trademark image generation and instance-discrimination-based feature learning is used in the trademark retrieval system.The weak ranking list of different models is integrated by the IRP fusion method to form a strong ranking list.The method considers the problem in practical application and has good computing performance and generalization ability.In addition,this paper proposes several optimization methods for the computational efficiency and memory occupancy problem of retrieval systems,and then designs an available trademark retrieval systems based on the retrieval algorithm proposed in this paper.This paper verifies and analyzes the proposed sampling algorithm on the image classification benchmark data set.At the same time,a large trademark data set containing922926 samples was used for trademark retrieval method training and testing.Finally the solution suitable for trademark retrieval was proposed with the experiment analysis.The experimental results show that both of the proposed methods can achieve satisfying performance and provide a new idea for deep embedding learning and unsupervised feature learning method design for trademark.