| As our country continues to pay attention to environmental issues,domestic waste classification has become a hot topic in today’s society,and more and more cities have begun to implement domestic waste classification policies.The relevant regulations on garbage classification divide household garbage into four categories: recyclables,hazardous garbage,wet garbage and dry garbage.If the recyclables are recycled and used reasonably,resources can be saved,while environmental pollution and damage can be avoided,and the concept of sustainable development in my country can be met.Therefore,the research on the classification and detection algorithm of recyclables has certain significance and value.In response to the above problems,the thesis has carried out work in the following aspects:First,establish an image library of recyclables in four categories,including plastic bottles,glass bottles,metal cans,and waste paper.The image library consists of a public image library and a self-collected image library.The public image library is provided by a third-party professional data platform Kesai.com,and 521 single-target data images have been screened.The self-collected image library is 100 self-collected images of multi-target recyclables,which are enlarged by 8 times by data enhancement technology,and contain a total of 800 data images.Label Img software is used to label the target category and location of the data image library,and after five-fold cross-validation,it is divided into training set,validation set and test set.Then,the thesis studies the existing target detection algorithms.Faster RCNN was selected as the classification and detection algorithm for recyclables,and three improvements were made to improve the detection accuracy and detection speed.First,Res Net50 is used instead of VGG16 as the feature extraction network.Then,according to the aspect ratio of the target in the data image database of the thesis,based on the K-means clustering algorithm,redefine the size of the anchor frame in the regional suggestion network.Finally,the global average pooling is used to replace the fully connected layer to complete the classification and regression work.In the model training stage,the weighted cross-entropy loss function is used as the multi-class loss function in the multi-task loss to reduce the misclassification of a small number of targets.Finally,experiment with the improved model on the Tensorflow deep learning framework.The experimental results show that the improved model of the feature extraction network improves the detection accuracy by 5.14%,due to the deepening of the model structure,the detection time has increased by 0.006 seconds per image,but the impact is negligible.After improving the model of the regional suggestion network,the detection accuracy is increased by 2.35%,the detection time is shortened by 0.012 seconds per image,and the average overlap is increased by 20.52%.After improving the classification and regression network model,the detection accuracy is increased by 1.62%,the detection time is shortened by 0.032 seconds per image,and the average overlap is increased by 1.65%.The overall improved model detection accuracy rate is 92.08%,the detection speed is 0.182 seconds per image,and the average overlap is 92.85%.Compared with the original model,the improved model performance is significantly improved and the robustness is stronger. |
