An Efficient And Lightweight Commodity Image Recognition Network

Unmanned vending machine is a hot field at present,in which the technology of unmanned vending machine after mechanization,electronic,digital gradually to intelligent development.Intelligent vending machine will use deep learning technology to realize functions such as speech recognition and vision-based commodity recognition.Among them,deep learning based commodity recognition technology is an important part in the landing application of intelligent vending machine.In the practical application scenario of vending machine,it is necessary to realize image recognition on an embedded device with limited computing power.This paper designs an efficient and lightweight commodity image recognition network—The Improved Mobile Net-SSD Network,on X86 platform and Android platform to realize the rapid and accurate identification.On the basis of the network,feature extraction network,feature matching and feature library are added,which to some extent reduces the time-consuming and laborious data marking and network training when adding new commodities.The main work and contributions of this paper are as follows:First of all,5000 images of five commodities under the scene of unmanned vending machine were randomly collected,among which the number of target commodities was 88,998.The commodity data set was made by marking the target goods manually,among which 4000 images were used as the training set for model training,and 1000 images were used as the test set for model testing.Secondly,On the basis of mobilenet-ssd network,the deconvolution and feature fusion structure are added and the convolution layer number is reduced.Then,the proportion and size of pre-selected boxes are adaptively adjusted to obtain The Improved Mobile Net-SSD Network.The models of The Mobile Net-SSD Network and The Improved Mobile Net-SSD Network were transplanted from Caffe environment to X86 platform for testing,and the average detection accuracy and detection speed of the model were: 0.919/44ms、0.948/33 ms.It can be seen that The Improved Mobile Net-SSD Network proposed in this paper has improved both the average detection accuracy and the detection speed.The model of the network with good detection effect was selected and transplanted to the Android platform for testing,and its detection speed reached 200 ms,proving the possibility of practical application of the network in commodity recognition tasks.Finally,When new products to be identified are added to the vending machine,a large number of images of new products need to be collected and manually marked,and the product image recognition network also needs to be retrained.In order to reduce the timeconsuming data marking and network training when adding new commodities,this paper takes The Improved Mobile Net-SSD Network as the detection network and adds feature extraction network,feature matching and feature library to design a new commodity image recognition algorithm.The commodity image recognition algorithm was transplanted to X86 platform for testing,and the average detection accuracy and detection speed of 1000 test images of five commodities were 0.974/237 ms.Without retraining the detection network and feature extraction network,different new products were added for testing in the unmanned vending machine environment,and the recognition of new products failed to achieve the expected effect.The reason is that the detection network based on five commodities has a poor ability to extract the features of commodities that are different from those of the five commodities,resulting in a poor detection effect;The feature extraction network does not pay attention to the distance between different commodity features,which makes it difficult to distinguish the new commodity features from the features of five commodities,resulting in poor feature matching effect.It shows that the algorithm of commodity image recognition proposed in this paper still needs to be improved.