Research And Implementation Of Commodity Detection Network Optimization Based On Automated Machine Learning

The retail industry plays an important role in a country’s economy.With the rise of artificial intelligence,unmanned retail develops rapidly.Among them,commodity detection is the key technology of unmanned retail,and it is also a hot topic in the field of deep learning.Deep learning technology relies heavily on expert experience and human intervention,while automated machine learning will automatically learn a series of steps such as model establishment and hyperparameter optimization for specific tasks,so that the entire process can get the optimal model without manual intervention.Therefore,this thesis will study the optimization problem of commodity detection network based on automated machine learning,in order to obtain the network with high detection accuracy and further optimize its architecture,thereby reduce the network complexity.The following is the main research content of this thesis:(1)Firstly,this thesis analyzes the current situation of object detection,and draws on the idea that the main network of most object detection algorithms is designed based on image classification.At the same time,it compares the neural architecture search methods based on automated machine learning,and selects ENAS method with high precision and full hardware support to neural architecture search for commodity classification task.Finally,the Faster RCNN algorithm is used to verify the effectiveness of the search network structure in the commodity detection task.(2)In order to solve the training bias and coupling problems of ENAS method,this thesis improves ENAS method based on creating temporary weights.By using hypernetworks to generate temporary weights for sub networks,the disadvantage of reusing all sub networks by using a set of weights is avoided.At the same time,the neural architecture search process not only needs to minimize the training loss,but also needs to maximize the verification accuracy.By applying the two gradients to the original weight and the temporary weight respectively,the two processes are decoupled,thereby reduce the training cost.(3)In order to solve the problems of limited performance and high computing cost of network architecture,this thesis proposes a network architecture optimization method based on NAT.For the network architecture with optimization space,a neural network architecture transformer is trained,which takes the network architecture as the input and the optimized network architecture as the output.Without introducing other calculation and parameters,the redundant calculation operation in the network architecture is replaced by the operation with less calculation and lower complexity,so as to obtain a more compact neural network model without affecting the accuracy of the network.The experiments show that it is feasible to search the commodity recognition network architecture based on automated machine learning,which can not only reduce the manual intervention in the process of model construction and training,but also further reduce the calculation cost by optimizing the network,thereby obtain a high precision and low complexity commodity recognition model and provide technical support for unmanned retail.