A Study Of Improved Few-shot Object Detection Method Based On Meta-learning For Scrap Identification And Classification

At present,the correct classification of waste products is puzzling the normal life of residents.The mainstream waste identification and classification system uses R-CNN f or identification,but the generalization ability of R-CNN is weak.Compared with diffic ult samples(such as occlusion,deformation,etc.),R-CNN can only improve the detecti on accuracy by increasing training samples.In addition to the long tail effect in real life,more training samples are required.Secondly,the research of intelligent waste recogniti on and classification system mainly focuses on the single structure design of waste box or small program.However,this lacks a composite structure that combines the two.(1)In view of the weak generalization ability of traditional R-CNN,this dissertatio n first introduces the traditional R-CNN algorithm.The meta learning method is used to t ransfer the meta knowledge learned from the data rich base classes to the new classes wi th scarce data.The coarse granularity prototype matching network is improved.The non linear classifier based on metric learning instead of the traditional linear object classifier is used to process the similarity between anchor and new class in the query image,thus improving the recall rate of new class candidate boxes with few shot.The fine-grained p rototype matching network is improved,and a module with spatial feature area matchin g and foreground concern is added to handle the similarity between noisy candidate box es and few shot new classes,so as to solve the spatial area mismatch between candidate box features and class prototypes,thus improving the overall measurement accuracy.In most sample indexes of The PASCAL VOC dataset,the improved algorithm in this diss ertation outperforms the previous most advanced algorithm by more than 3.0m AP50.Si milar precision improvement is achieved on MSCOCO datasets.Secondly,a few shot de tector is designed.The classifier based on softmax and the few shot detector designed in this dissertation are considered together.Taking advantage of the advantages of these t wo detectors,the feature backbone network is shared by using the few shot detector in t his dissertation,and a Faster R-CNN detector head is jointly learned to detect base class es and new classes.On the basis of maintaining the original detection accuracy,the dete ction range is expanded.(2)This dissertation combines the software and hardware of applet and waste box.The software part is based on We Chat applet,including a applet side and a server side.T he applet side is responsible for displaying the front-end interface,while the server side is mainly responsible for interacting with the applet side.This interaction has the follow ing functions: system homepage,image recognition,text recognition and integral test.T he hardware part is based on raspberry pie,and is divided into raspberry pie and waste b ox itself.The former acts as the "brain" of the latter,deploys the improved algorithm mo del in this dissertation to raspberry pie,and is responsible for identifying,classifying an d loading waste products.In addition,it can also complete image capture and other func tions,and propose waste classification suggestions to users through the automatic broad cast function.Figure 24 table 14 reference 77...