Research On Classification Algorithm Of Recycled Plastic Bottles Based On Faster R-CNN

In our life,the vast majority of food and articles are packed with plastic products.For example,almost all drinks are sold with plastic bottles.A large number of plastic bottles not only bring us convenience,but also cause white pollution.The recycling of plastic bottles can greatly reduce the damage to the environment,However,at present,the waste recycling work is mainly done manually.Due to the lack of manpower,large quantities of waste plastic bottles cannot be recycled and classified in time.However,the recognition accuracy of traditional machine vision algorithm for waste plastic bottles with complex and diverse shapes is not ideal.In order to improve the recycling efficiency and classification accuracy of plastic bottles,This paper presents a high-precision image classification algorithm for recycled plastic bottles based on faster r-cnn.Aiming at the problem of low classification accuracy caused by uneven distribution of recycled plastic bottle data,this paper proposes a joint prediction method of fast r-cnn in deep learning and RESNET model,which solves the problem that the accuracy of single target detection model can not meet the recovery requirements in the case of unbalanced categories.At the same time,soft NMS(flexible non maximum suppression)is used to improve the detection accuracy of overlapping plastic bottles.Further using ohem algorithm,the algorithm proposes to use two same ROI networks to filter out difficult negative samples,which balances the class imbalance problem between difficult samples and simple samples in the training process of fast r-cnn.Through the first ROI detector responsible for forward propagation,the simple samples with smaller loss value are filtered out.At the same time,in order to further improve the detection accuracy of faster r-cnn model,the ciou loss function with the proportion of prediction and true value rectangular box is used to improve the regression accuracy of prediction box.In order to solve the problem that there are too few samples of some categories in the color classification data set of recycled plastic bottles,the model can not learn enough sample features,and the classification accuracy can not be improved,an image expansion software is developed by using pyqt to supplement the images with too little data and narrow the data imbalance gap,At the same time,in the training process,the weight sampling method is used to set the sampling probability of each class,so that the image data of each batch can be sampled to all classes of samples,which further improves the problem of sample imbalance and further improves the accuracy of the classifier.Finally,the experimental verification of the proposed algorithm improvement method is carried out,and the optimal improvement strategy is obtained through the analysis of the prediction results of the algorithm model obtained by different improvement methods.The accuracy and speed of the model are tested.The method meets the requirements of classification accuracy and speed.Finally,we propose to deploy the algorithm to the web side by using the flash framework to visualize the actual prediction results...