The Research On Logo Recognition Based On CNN Tree

With the development of society and the continuous improvement of social informatization,content-based image retrieval has become a popular research direction in today.Among them,logo recognition has become a very important research direction due to its broad application fields and huge commercial value.Although more and more scholars have invested in the research of logo recognition,various logo recognition techniques are also numerous,but because real-life logo are easily affected by the filming angle and the external environment,they bring various changes to logo images,such as scale changes,rotation changes,affine deformation,and partial occlusion variations.All these problems make the logo recognition algorithms face great challenges.Currently,the relevant algorithms of logo recognition in natural scenes can not meet the actual application requirements in accuracy and speed.In recent years,the methods of deep learning have achieved breakthrough results in the areas of image classification,object detection,and object segmentation.This dissertation studies the logo recognition algorithm in natural scenes,using the related technologies of deep learning combined with the idea of decision tree and spectral clustering algorithm.The main works and contributions of this dissertation are provided as follows:(1)A logo recognition algorithm based on CNN Tree is proposed.At present,most logo recognition algorithms based on the convolutional neural network only use a single CNN to build the model.For recognizing more complex categories,the recognition accuracy needs to be improved.This dissertation proposes a logo recognition algorithm based on CNN-T,which makes full use of the recognition performance of CNNs in the multiple layered dendrogram recognition model.Each CNN is used as a multi-classifier to identify a specific cluster which is generated by spectral clustering algorithm.The algorithm proposed by this dissertation makes full use of the classification performance of multiple CNNs and improves the accuracy of logo recognition.The experimental results show that the logo recognition algorithm based on CNN Tree has achieved remarkable results.(2)A spectral clustering algorithm based on feature fusion is proposed.In this dissertation,the idea of feature fusion is integrated into the spectral clustering algorithm and the characteristic that shallow convolutional neural network and deep convolutional neural network can learn different features of the object is used,and this dissertation proposes a spectral clustering algorithm based on feature fusion.The algorithm proposed by this dissertation fuses the shallow characteristic fea tures of the network with the deep semantic features of the network in a specific way.The algorithm improves the accuracy of the logo clustering,and then improves the recognition accuracy of the entire CNN-T system.The experimental results show that the spectral clustering algorithm based on feature fusion can significantly improve the accuracy of logo recognition.