Research On Automatic Identification Method Of Conveyor Belt Sushi

The automatic identification of the conveyor belt sushi is a blank field.Method using RFID can only identify the type of the plate, not the type of thesuhsi.There are two actual problems according to the sushi restaurant operators.The first one is that sushi may exit too long to maintain freshness and taste,which will certainly reduce the quality of the restaurant. The second one isthat there are leftovers and waste since the runner cannot catch the customers’needs dynamically.The automatic identification method is an indispensable part in theautomated management of the restaurant. A recognition algorithm which runsstably and efficiently can help the manager make a precise estimation over thetype and the quantity, have a better understand of customers’ needs, and makethe right decision, enhance the economic efficiency and quality control, andavoid unnecessary wastes.Sushi identification is a specific area which belongs to the imageidentification area. So some common methods in image identification can alsobe applied in sushi identification. However, sushi identification also showsparticularity in some specific aspects. First, the direction sushi been placed israndom. Second, the system needs to run all-day long and all-weather.This paper has made a research on the automatic identification method ofthe conveyor belt sushi images based on the characteristics, the direction ofsushi is random and the illumination intensity varies greatly, above andrelative experiments.Firstly, the paper extracts the color features including the HSV globalhistogram, color moments, and the HSV accumulative histogram with theconsideration of the random-direction chacrateristics. The paper alsoimproved the HSV accumulative histogram, which makes the accuracy of thesushi identification higher.Secondly, the paper extracts the texture features using the GLCM andLBP\LBPC\LBPV descriptors as an supplementary information of the colorfeatures. The texture features also contributes to the robustness of the changesof illumination.At the end of the paper, a few comparative experiments on featureextraction algorithm has been under discussion first, then experiment on thefusion weights optimization, the experiment of scale selection and differentclassification methods have been shown including the minimum distance classifier, the KNN classifier and SVM. Finally, an efficient scheme with anideal accuracy and acceptable efficiency has been proposed.Experiments have shown that the method has good classificationaccuracy, which is higher than the required90%.