Research On Commodity Detection Algorithm Based On Deep Learning Under The Background Of "New Retail"

Retail industry has always been an important part of our national economy.In recent years,due to the development of online transactions,offline retail transactions are shrinking.However,with the introduction of the new retail concept,online development has encountered a bottleneck period,and more and more enterprises are re-focusing on offline development.The traditional settlement method requires a lot of labor,which is inefficient and the cost is gradually rising.New technologies are needed to improve the existing bar code settlement method.Deep learning algorithms have developed rapidly since 2012,and image classification and target detection algorithms based on deep learning have also been further developed.In this context,this thesis takes commodity settlement in the context of new retail as an entry point,studies commodity detection algorithms based on deep learning,uses fresh bakery as a transaction scenario,and optimizes some of the detection algorithms to improve detection effect.This thesis focuses on a convolution neural network model based on Faster R-CNN algorithm,which improves the detection accuracy compared with the original Faster R-CNN algorithm.The main work of this thesis is as follows:(1)Firstly,the basic network is tested,VGG16 is replaced by ResNet network with higher detection accuracy,and ResNet-50 is selected as the basic network considering the practical factors.Then,inspired by GoogLeNet Inception,the ResNet network is optimized.The original convolution layer feature extraction effect is limited.We hope to increase the richness of receptive fields in the residual network and design different receptive field modules.In order to simplify the calculation,7×7 is replaced by 7×1 and 1×7 cascades.The combination of these modules was tested and the combination with moderate precision and speed was selected.(2)Considering that customers often choose bread of different sizes in real settlement,it is difficult to detect small targets.In order to further improve the detection accuracy of small targets,the feature pyramid is applied to the settlement algorithm model.(3)Then,the improved algorithm is implemented and different experimental schemes are designed to evaluate.The experimental results show that the new algorithm performs well on our data set.Compared with the algorithms based on simple GoogLeNet v3,ResNet-50 and ResNet-101 on Recall rate,Precision rate and mAP,the new algorithm has a certain improvement,which can explain the significance of the application of the model in actual scenes.