The Research On Logo Detection Based On Image Classification And Object Localization

In the information service platform for printing image security application,first of all,we need to carry out trademark detection operation based on image classification and target location for the commodity packaging image uploaded by the user,then obtain the standard image through image correction and other processing steps,and then transmit the standard image to the cloud to analyze the security information.Image classification and target location technology is the key technology of trademark detection service.The speed and accuracy of classification and location directly affect the user experience.Due to the variety of trademark styles,the complexity of background and the influence of many factors such as illumination,occlusion and blur,the opportunities and challenges of trademark detection algorithm research coexist.The main contents of this paper are as follows:1)This paper constructs 83 kinds of trademark image classification and recognition data set DDlogos,selects VGGnet16 and Res Net50 to train and study the user-defined trademark data set DDlogos,and tests the average accuracy of 83 kinds of trademark data using VGGnet16 network is 98.24%,using Res Net50 network is 99.91%.2)This paper constructs 24 kinds of trademark image target detection data set DDlogosOD.In the target detection algorithm based on candidate region,Faster R-CNN and its extended network Mask R-CNN are selected to train and learn the customized trademark data set DDlogos-OD.Through the analysis of test results,it is found that the detection accuracy and speed of Mask R-CNN are better than that of Faster R-CNN.Then SSD and YOLO series algorithm are selected to train and learn the user-defined trademark data set ddlogos-od in the target detection algorithm based on regression method.The test results show that YOLOv3 has the highest detection accuracy and its detection speed is twice that of Mask R-CNN.3)This paper summarizes the data set construction methods used in the network structure of image classification and target location,forms a tool for making standard data set,and constructs a trademark detection system based on hybrid network,introduces the framework of a complete trademark detection system,which uses Res Net50 for the task of trademark image classification and recognition,and uses Mask R-CNN and YOLOv3 to complete the trademark target detection Task.The research content of this paper has achieved good results in the practical engineering application of trademark detection service.