Research On Garbage Classification And Recognition Algorithm Based On Deep Learning

With the rapid economical development of our country and the rising of the living standards of the people,how to deal with household waste has become an important research subject.Therefore,our country has issued relevant laws and regulations for the classified recycling of household waste.At present,some cities have issued household waste management regulations for providing guidance to the public.In order to promote the reform of traditional household waste collection and improve the efficiency of garbage sorting,this paper puts forward an improved object detection algorithm based on deep learning,which is applied to garbage classification and recognition to make garbage classification more intelligent.According to these national policies about garbage classification at the present stage,this paper collects pictures of 40 different types of household garbage as a data set based on the classification criteria of recyclables,kitchen waste,harmful waste and other garbage.In order to reduces the gap between unbalanced categories of data set,image enhancement is adopted to expand the data samples.Based on Faster R-CNN,this paper improves the structure of feature extraction network to solve the problem of poor recognition in small target objects.The feature maps generated by the underlying convolutional layer has higher definition and it extracts spatial properties such as edge,texture and color information,which is conducive to detect the small objects in the image.The feature maps generated by the upper convolutional layer has larger receptive field and it is more abstract,which is conducive to detect the large objects in the image.Therefore,feature fusion is used to improve the precision of small object in this paper.Besides,multi-size convolution kernel is used to enhance the generalization of the network for identifying target objects of different sizes in the RPN.Simultaneously,according to the characteristics of experimental data sets,this paper has designed the size and scale of anchor boxes in the RPN to improve the accurate of target detection boxes.Compared to Faster R-CNN,the improved algorithm promote the m AP in the test set,without significant decrease in detection speed.