Research On Data Generation And Object Recognition Algorithms Under Unmanned Retail Environment

With the rapid development of technologies such as mobile payment,deep learning,and cloud computing,the transformation of traditional retail models has become possible.Research on deep learning algorithm has become a hot direction in computer vision and digital image processing,and deep learning techniques are widely used in face recognition,intelligent transport system,smart city and many other fields.In the concept of “unmanned retail”,deep learning algorithms such as object detection and image classification play a central role.At present,smart unmanned vending machines are the representative carrier of unmanned retail business models.Compared with traditional vending machines,they have greatly improved user experience and implementation technology.The one core issue of smart unmanned vending machines is how to quickly and accurately identify which and what kind of items the user is taking.In this paper,deep learning is used to solve the problem of object recognition in smart unmanned vending machines.To accurately obtain the information of users' goods,a dynamic object recognition method based on image classification and a static object recognition method based on target detection are proposed.However,open-source datasets for classification and object detection have not been optimized for unmanned retail applications,which cannot be used directly to train models.Therefore,to promote the development of unmanned retail by using deep learning-based classification and object detection,this thesis builds an experimental platform and established a dataset,which contains 34052 images collected from smart unmanned vending machines scenarios.On this dataset,the mainstream image classification and object detection model are trained,and performance evaluation is conducted,and an object detection model,namely Drt Net,was proposed and designed for the smart unmanned vending machines scenario.The backbone network of the model uses deformable convolution and group normalization layers,and the loss function is focal loss function and balance L1 loss function,which improved the recall rate of beverage detection.The feasibility of applying the deep learning model to smart unmanned vending machines is verified by experiments.Besides,since a large amount of annotation data is required for training a deep learning model,and there are many kinds of commodities in the retail field,it brings great challenges to data collection and labelling,resulting in a lack of labeled data.Traditional methods solve the lack of labeled data.Generally,data augmentation methods such as flipping,translation,scaling,rotation,and random cropping of existing image data are used,but such methods lack adaptability to new data types.In this thesis,a data generation algorithm by leveraging a deep CNN-based encoder-decoder model for saliency detection and generative adversarial network for image-to-image translation is proposed,which can automatically generate labeled data.The algorithm uses the improved Uper Net50 network for saliency object detection,and Cycle GAN for image style transfer.Finally,the generated image is used as the training set of the object detection model.The test is performed on a public unmanned retail data set,namely RPC,and the test results verify the effectiveness of the proposed algorithm.