| Since the 21 st century,the application of image recognition technology has become more and more extensive,and the rise of deep learning has made image recognition a new leap-forward development.The traditional target recognition method has considerable limitations.When dealing with the obtained data set,complex pre-processing of the image is required.When processing image features,when extracting the features of the image map,it is necessary to combine the specific features of the image to design the corresponding Methods.The classifiers constructed by some traditional extraction methods have low recognition accuracy and are often only applicable to specific objects.In this study,we first elaborated on the limitations of traditional methods,and then developed a new image recognition technology based on deep learning,which uses two types of convolutional neural networks.The feature learning function of this kind of network is relatively powerful,which can extract image information and improve the detection performance.The research object of this paper is the RMB watermark.By building the RMB image data set,the experimental results of the classic target detection network are verified,and the deep learning model is developed and optimized to accurately identify the RMB watermark.In the research process,the scale-invariant network and the target detection algorithm Refine Det were optimized.Based on experiments,the performance of these two algorithms in the field of watermark recognition was confirmed,and their performance compared with the classic target detection network was compared.Application effect.The main work of this study includes:Expand the RMB image information.In view of the current lack of a suitable open RMB image data set,a variety of RMB images in the actual natural environment were manually collected and annotated.The RMB image data set is expanded using various methods such as image rotation and image noise addition.After the expansion process,the number of sample images in the data set is up to 4,000,which has been greatly improved,which is enough for deep learning and training.Develop a deep learning framework for identification in this field,and then compare and verify it through corresponding experiments.In this paper,based on the research characteristics of RMB watermark recognition,on the basis of existing research,the target recognition algorithm is analyzed and selected.In the comparative experiment,classic target detection algorithms such as Faster-RCNN and SSD are selected as the main research framework of this experiment..Based on the analysis and discussion of RMB images,the optimized Refine Det and scale-invariant networks aregiven,and they are applied accordingly.Experimental results show that the improved network can improve the real-time recognition of RMB image data sets.This paper summarizes and analyzes the deficiencies of Faster-RCNN(VGG19)model applied to RMB image data sets.When selecting the target detection algorithm,the optimized algorithm is proposed and analyzed.The experiment verified the performance of the improved scheme.At the end of the thesis,the proposed algorithm is experimentally verified.From the results,it can be seen that the method proposed in this paper has good adaptability and real-time performance. |
