Solid Metal Garbage Classification And Detection Based On Deep Learning

Garbage is a misplaced resource,compared with household waste classification and recycling,solid metal waste not only has a huge recycling space but also has great economic reuse value.However,in our daily life,the classification and recycling of metal waste still rely on the traditional manual classification,which is time-consuming and laborious.In view of this actual situation,this paper proposes a method based on deep learning to complete the classification and detection of metal waste,and mainly completes the following two parts.(1)For the classification of solid metal waste classification.First,a total of 17,804 metal waste images of six classes,including battery,can box,CD player,metal bowl,soda can,protable battery.,were collected and named as gx-trashnet.Second,the classification of performance of VGGNet,Res Net,Inception-V3,Mobile Net and Dense Net on Trashnet datasets and gx-trashnet datasets is compared respectively by using transfer learning method.The SE attention mechanism module was added into the Res Net-101＿V1 with the best classification performance.The full connection layer of the SE attention mechanism module was replaced by a convolution layer with a convolution kernel size of 1×1,and the global average pooling layer was used as the classifier.Through experimental demonstration,the classification accuracy of the improved network model achieves 97%accuracy on gx-trashnet.(2)For the detection of solid metal waste,.First the constructed data set was annotated.Second,in order to determine the target detection algorithm more suitable for this data set,SSD＿300 and YOLO V4 models were compared,and YOLO V4 with better detection performance was selected as the metal waste detection framework.Because the number of Res Net-101＿V1 parameters proposed in the classification task is too large,in order to balance the detection speed and detection accuracy,the lightweight Mobile Net is used as the trunk network of feature extraction.Before the experiment,the K-means clustering algorithm is used to adjust the size of the anchor boxes to make the detection network more suitable for this data set.Compared with the original research results,the experimental results reduced the number of parameters by 10 times,and obtained 96.63%m AP,96.84% Precision,90.68 Recall and 0.94 F1-Score on gx-trashnet,tested each image average use 32.855 ms and got 30.389 FPS.