Research On Recyclable Garbage Detection Based On SSD Algorithm

Recyclable garbage is a kind of domestic garbage suitable for recycling and resource utilization.Recycling and processing garbage can reduce pollution and save resources.There are many types of recyclable garbage.The classification efficiency of artificial garbage is low and easy to be misclassified,and the recycling rate of garbage is not high.In order to improve the efficiency of garbage classification and make full use of garbage resources,a garbage detection model based on Single Shot Multibox Detector algorithm is designed for automatic detection of recyclable garbage.In view of the characteristics of One-step object detection method in image detection,this thesis has done the following four aspects for some key technologies and theoretical methods:(1)According to the definition of recyclable garbage,a dataset for recyclable garbage detection is constructed.Two main object detection algorithms and their application scenarios are studied.The One-step object detection algorithm is selected to build a recyclable garbage detection model and detect the recyclable garbage dataset.The detection results verify that the network is suitable for recyclable garbage dataset.(2)By comparing the classification effects of different types of classification models on recyclable garbage dataset,ResNet is selected as the feature extraction network of SSD.Aiming at the possible overfitting problem of deep learning model,a model preventing overfitting scheme is designed from three aspects of loss function optimization,batch normalization and data expansion to improve the generalization ability of the detection model.Aiming at the problem of low accuracy of small object detection in object detection network,two data preprocessing schemes are designed to improve the accuracy of small object detection.Experiments show that residual learning can help the gradient to be transmitted effectively.The improved SSD network improves the detection accuracy and maintains a high detection speed.(3)Based on SSD network structure,the influence of feature fusion scheme and sample equalization on model performance is analyzed,and an optimized recyclable garbage detection network Filter-SSD is designed.The Link Block is added to the network,and the connection between the high-level features and the low-level features is established through the feature fusion scheme to reduce the loss of semantic information in the down-sampling process of the feature map.The Filter Block is added before the prior box classification and regression of the model to balance the positive and negative samples to improve the problems caused by sample imbalance.The experimental results show that the optimized Filter-SSD has higher detection accuracy and lower missed detection rate.(4)An optimization model O-SSD is designed based on Filter-SSD network.Octave Convolution is introduced to replace the traditional convolution in the model,which reduces the parameters and computation of the whole network.The feature pyramid structure is extended,and the prior box is extracted from the lower-level feature map to adapt the small object and increase the matching degree,which improves the accuracy of small object detection.A Res-Inception feature extraction module is designed to enrich the feature quantity of shallow feature map from two aspects of network depth and network width.Through comparative experiments,it is verified that O-SSD has higher detection accuracy and computational efficiency than Filter-SSD,and has better performance on recyclable garbage dataset.Through the performance comparison and evaluation with the existing algorithms,it is proved that the O-SSD algorithm proposed in this thesis can effectively improve the detection accuracy of recyclable garbage.It has less computation under the complex model structure,better balances the accuracy and speed of the model,and has better actual detection effect on recyclable garbage.