Research On Item Identification Based On Neural Network

With the rapid development of Internet of Things technology and China's rapid development of smart home appliance business,smart refrigerator as a representative of intelligent white,has become a breakthrough in the development of enterprises competing for the product.Among them,food management is one of the core functions of smart refrigerators,and also is the key to the identification of items.In this thesis,a multi-target detection algorithm based on neural network is introduced for the identification of refrigerator items.First of all,this thesis gives a brief introduction to the application background of the product identification technology utilized in intelligent refrigerator and the present situation of corresponding domestic and foreign development.Moreover,the thesis briefly analyzes some existing problems and introduces some basic related knowledge,which provides a theoretical basis for the study of object recognition algorithms based on CNN.Secondly,traditional image recognition algorithm on items has many drawbacks,including complex calculations,time overhead and low accuracy when adopting features of artificial design.Aiming at these drawbacks,this thesis proposes a new method which applies CNN to extract the depth characteristics of images instead of features of artificial design.The new method can overcome the disadvantages of artificial extraction and realize the automatic extraction of the characteristics of images.Because the image of the refrigerator object usually has many kinds of targets at the same time,and the convolution neural network is the classification of the whole image,this thesis introduces the multi-target detection algorithm based on the convolution neural network,namely the Faster Region with Convolutional Neural Network(Faster R-CNN),and applies it to the refrigerator item identification task,to achieve the identification and classification of items in the refrigerator.On the basis of this,this thesis makes an effective optimization based on the characteristics of the actual scene of the refrigerator,thus improving the detection accuracy of the item recognition algorithm.Because Faster R-CNN is based on convolution neural network,this thesis verifies the influence of different CNN models on the accuracy of Faster R-CNN recognition by changing different convolution neural network models.At the same time,A light source experiment verifies the robustness of the algorithm.Finally,it is found that the Faster R-CNN appears some problems at the time of detection that are concluded as: the confidence threshold is increased,the missed detection occurs,the confidence threshold is lowered and the false alarm occurs.Region Proposal Networks(RPN)of Faster R-CNN is highly accurate when it detects whether there is a target,so it can take advantage of RPN to solve the problem of missed and false.Therefore,this thesis proposes a re-recognition method,which combines the re-recognition and the optimized Faster R-CNN algorithm to solve the problem of large part of the missed and false.Experiments show that the improved Faster R-CNN algorithm has the optimal performance in the refrigerator scene environment.