Logo Recognition Method Based On Target Detection And Image Retrieval Technology

With the economic growth,the modern business center has gradually developed into a large-scale commercial complex integrating shopping,dining,socializing and entertainment.Because large commercial complexes are multi-storey buildings,GPS cannot provide reliable location information and navigation services,and various store logos in commercial complexes have become an important reference for location information.The device uses visual technology to perceive and identify the business' s logo information in the business environment,which plays an important role in positioning and navigating the user's location.When detecting and recognizing store logos in natural business scenes,the detection and recognition algorithms for store logos are often affected by many types of brands,complex and changing environmental backgrounds,different lighting conditions,randomly generated occlusion or blurring,etc.It is a challenge that must be faced.At present,existing store logo recognition algorithms have not yet been able to meet the requirements for large-scale store logo detection and recognition in natural business scenarios.Thanks to the development of chip technology,deep learning technology has made great progress under the support of massively parallel computing chips.Convolutional neural network is a deep learning technology that is very suitable for processing image data.This article proposes a method of combining target detection and image retrieval through convolutional neural network to achieve the purpose of identifying store logos in natural scenes It finally laid the foundation for shopping mall positioning and visual navigation services.At the same time,it proposed the method and process of deploying the algorithm on the mobile terminal and the server and implementing the deep learning algorithm on the specific product.The main work includes:1.Propose the realization scheme of Logo recognition algorithm,first determine the position of Logo in the image through target detection,and then encode the Logo area in the image and search with the Logo feature code in the library,and use the retrieved Logo label as The label of the logo to be recognized.2.Produced an image data set containing more than 4,000 shopping mall and street scenes,of which more than 2,700 were obtained by independent shooting and 1300 were obtained by Internet crawlers.3.Drawing on the ideas of One-Stage and Two-Stage target detection algorithms and the attention mechanism,the existing target detection algorithms are improved,and an efficient logo detection method based on the attention mechanism is proposed.4.In order to be able to run smoothly on the mobile terminal,draw on the ideas of the Mobile Net V2 network,reduce the calculation parameters of the model through deep separable convolution,and use the channel pruning method to prune the convolutional neural network model weights.5.Research and compare the advantages and disadvantages of different mobile deep learning inference frameworks,select the appropriate mobile inference framework MNN to deploy the Logo detection model.6.Express the features of the Logo image based on the convolutional neural network and use the principal component analysis method to reduce the dimensionality and compression of the features output by the convolutional neural network.7.Deploy the Logo retrieval algorithm to the server via docker.