The Research On Logo Detection And Recognition Based On Deep Learning

As people enter the era of mobile Internate,the rapid development of location information services has gradually changed people's daily lives,providing people with more convenient,more abundant peripheral information services.Location technology is one of the key technologies of location information services.Because the intensity of satellite signal is weakened by building block,it will result in inaccurate positioning of the indoor positioning system.In order to solve the problem of inaccurate indoor positioning,this dissertation uses the location of logo to improve the accuracy of indoor positioning From the the perspective of pattern recognition,we can use logo detection and recognition algorithms to detect and identify the logo in indoor shopping mall,so as to improve the accuracy of indoor location.In the natural scene,because of the variety of the trademark,the complexity of background and the interference of the illumination,occlusion and blur,it is a great challenge to the logo detection and recognition algorithms.At present,the algorithm of logo detection and recognition in natural scene can not meet the needs of practical application in terms of accuracy and speed.In recent years,the methods of deep learning have achieved breakthrough in the research areas of image classification,object detection and image segmentation.This dissertation studies the logo detection and recognition in nature scene by the technology of deep learning.The main works and contributions of this dissertation are provided as follows:1.A logo dataset for indoor shopping environment is proposed.This dataset has 142 categories and 10920 manually labeled logo images,is the most categories for the currently published logo datasets.It contains a wide range of trademark styles,complex background and a variety of lighting,occlusion,halo and other interference factors.It is a great challenge to the logo detection and recognition algorithms.2.Logo detection algorithms based on attention and regression are proposed.At present,the detection algorithm based on candidate proposals becomes the mainstream algorithm to realize the logo detection using the deep learning technology.However,the candidate proposals-based logo detection algorithm is numerous and time-consuming.In this dissertation,based on the speed and precision of logo detection,a new algorithm based on regression is designed.However,because of too many negative samples in the detection results,the detection accuracy is low.In order to solve this problem,this dissertation proposes an algorithm based on attention and regression.The algorithm uses a branch of the network to extract the logo mask to provide additional supervision information to improve the precision of logo detection.Both qualitative and quantitative results confirm that the logo detection algorithm based on attention and regression can achieve remarkable performance.3.A new algorithm for logo recognition based on mixed feature is proposed.At present,most of the features used in the logo recognition are the underlying features and the depth features of the network training on the general objects.However,these features are limited to expression of the logo.This dissertation make full use of the logo with similar features of the text,a convolution neural network is proposed to extract the features of the logo from the perspective of text recognition.Otherwise,the logo is more unique than the text,such as the color,shape and other characteristics.In this dissertation,we use the shallow and deep layer of convolution networks to learn different characteristics.A hybrid feature recognition algorithm is proposed.This algorithm improves the accuracy of the logo recognition by fusing the underlying features of the shallow layer and semantic features of the deep layer in different ways.Experiments validate that the fusion method of hidden layers across together has a significant improvement on the logo recognition.