Research On Intelligent Visual Classification Method Of Waste Plastics Base On Machine Learning

With the rapid development of economic level and the growth of population,the renewal speed of domestic plastic products and the generation of more demand have been accelerated.Meanwhile,the output of waste plastic is rising rapidly,and its categories is becoming more diverse and complex,which brings great commercial potential and production pressure for the rough classification of plastic.With the rapid development of machine learning intelligent algorithm theory,the detection technology which takes visual camera as sensor and intelligent algorithm as the core of processing is gradually applied to various scenes of industrial production,which improves automation and reduces production cost.This paper studies the complex detection scene in the visual separation of waste plastics,and the main research contents are as follows:Firstly,the paper studies the classification method of waste plastics based on traditional machine learning SVM classifier,which includes three main processes: traditional image processing,feature extraction and SVM classification.Traditional image processing includes image filtering and threshold segmentation.Based on the traditional SVM algorithm,the multi classification model of SVM is designed for 7 categories of waste plastics in this study,and the corresponding experimental tests are carried out.Secondly,the detection method of waste plastics based on convolutional neural network is studied,which includes data preparation,model training,generalization test and data augmentation.This paper mainly improves the Adan optimizer,and trains three detection frameworks based on anchor base and anchor Free: SSD,Center Net and Foveabox.The result shows that Foveabox is established as the detection framework of this study by comparing experiments.Finally,according to the shortcoming of Foveabox in waste plastic separation scene,and the complexity of the waste plastic target,this study is focused on algorithm improvement.Firstly,in order to improve the effective sense field of convolution process,a deformable convolution technology with scaling coefficient is proposed for the complex characteristics of the waste plastic target.Secondly,in order to improve the detection accuracy of the near target,a Soft-Weighted-Anchor mechanism with hierarchical control factor is proposed to improve the detection accuracy of the target near the distance;thirdly,aiming at improving the effective sense field of convolution process,A multi-layer feature fusion scheme based on circular pyramid is proposed to replace the original feature pyramid.And each improvement scheme is verified by experiments.A series of experiments are designed to compare the detection effect of the improved Foveabox and other detection frameworks in the waste plastic separation scene,and verify its effectiveness and feasibility.In conclusion,the improved algorithm has better performance in the waste plastic classification scene,which meets the requirements of industrial production accuracy and real-time production,and has better migration ability and practical value.