Kitchen Utensils And Appliances Accessories Recognition And Sorting Based On Deep Learning Classification Algorithm

With the rapid change of the Internet,deep learning as an important branch of machine learning is increasingly popular.To enable computers to think like humans,humans have long explored the intricate nervous systems of animals,which have been realized and served by computers in human life.The mode of deep learning simulating neurons to transmit and process information has been widely used in image,speech,video,text and data analysis due to its accuracy and efficiency.The first application field is image recognition,which has been well applied in face recognition and license plate recognition.However,in industrial production,there are a large number of simple and repetitive human work,which can be completely replaced by computers.Moreover,difficulty in recruitment and tight order time have become the most fatal problem for the development of the factory,so the transformation and upgrading of the factory are extremely urgent.The computer can work 24 hours without a break,which not only speeds up the work efficiency and shortens the order completion time,but also replaces the manpower to reduce the manpower cost.It is a project with long-term return on investment for the enterprise.China,the author of this paper,double horse,LTD’s biggest-selling product,multi-function large shredder demand but the present situation of the limited manpower,multi-function shredder baotou cover and automated sorting work of science and technology research,the core work is in order to speed up the assembly line production efficiency and reduce labor costs training out of the automatic identification of the shredder in baotou and middle model,nature for image recognition classification problem.Firstly,200 samples of 7 colors of the front cover and the front and back sides of the middle body of the dish cutter were collected from the field in the assembly line,with a total of 4,200 samples.After batch naming of the collected images,data enhancement was carried out to expand the sample number to 20,000.Finally,Label Img was used to mark the images to generate XML files.Secondly,the tensorflow deep learning framework was built on the Linux system of Intel i7 processor 32 G memory and graphics card NVIDIA Ge Force RTX2080 Ti.Thirdly,70% of the data were randomly selected as the training set,and the rest were the test set.The product sample data were directly used to conduct training on the four models of faster_rcnn_inception_v2,faster_rcnn_resnet 50,ssd_mobilenet_v2 and ssd_inception_v2.After training 200,000 steps in each model,the accuracy of faster_rcnn_inception_v2 model reaches 93.89%,but this model has high requirements on hardware.Fourth,the use of the migration study above training model,found that in addition to faster_rcnn_resnet50 model of other model accuracy is improved obviously,ssd_inception_v2 + migration learning model accuracy is relatively large,the precision of the study with faster_rcnn_inception_v2 + migration model,but the SSD model for hardware requirements is low,and error recognition only occurs within the categories of small class,between the smaller error identification cost.Finally,these models were used for product identification and sorting,and ssd_inception_v2 + migration learning model was selected after comprehensive consideration.First,in terms of performance measurement,the accuracy of ssd_inception_v2 + migration learning model is only 0.25% different from that of the model with the highest accuracy.Second,in terms of running speed,compared with other algorithms,SSD series algorithm is the fastest.Thirdly,in terms of equipment application,considering that the final training model will be put on the raspberry pie for calculation,it is more appropriate to choose the SSD series model that can run smoothly on the phone.Fourth,in terms of error identification cost,according to the actual production situation,error identification occurs between small classes within a large class with the lowest cost,while error identification of ssd_inception_v2 + migration learning model only occurs between small classes.