Research On The Household Garbage Image Recognition Model Based On Knowledge Distillation

This paper mainly based on the related methods of deep learning in the field of computer vision,to carry out research on the recognition of household garbage.household garbage.Household garbage Household garbage is the product of human society’s metabolism,and since the birth of human society,the output of household garbage has risen with the development of society.In recent years,the environmental problems caused by the rapid increase of waste production are particularly prominent.Garbage classification correctly can not only reduce the negative impact of garbage on the environment,but also bring considerable resource value and economic value.The algorithms based on deep learning are used for image recognition,which have the advantages of low price,easy maintenance,reusable and strong stability.Therefore,with its unique advantages in the image field,deep learning can be used to solve the problem of garbage classification and has good prospects.Considering that the current deep convolution model maintains a high accuracy,but the number of parameters of the model is very large,which brings great difficulties to the deployment of the model on the machine.This paper proposes a method of household garbage identification based on knowledge distillation,which can improve the accuracy of household garbage image recognition without increasing the amount of model parameters.It is helpful to the field of household garbage classification and improve the efficiency of household garbage classification.The household garbage image recognition model proposed in this paper can achieve the characteristics of fast recognition,easy to use and convenient maintenance.This model can not only be transplanted to mobile phone software of household garbage classification,but also be applied to smart dustbins,garbage classification factories and other places.The main research contents of this paper are as follows:(1)Select the common household garbage from these four categories to build the data set.By simulating the real garbage state,images are collected from different backgrounds,angles,illumination and resolution.For some garbage with a small quantity,automatic amplification is carried out by various methods.The dataset includes 16,218 images and 30 garbage categories.The current popular object detection algorithms are used to evaluate the garbage data set.Finally,the data set is tested and analyzed,which proves that the data set can be used to evaluate household garbage recognition algorithms and has research significance.(2)By migrating the underlying features of natural images and applying them to garbage detection.Experiments on object detection model SSD show that this method can not only speed up the training process,but also improve the accuracy of the model.The Mobilenet network shines because of the introduction of depthwise separable convolutions.Inspired by the depthwise separable convolution in Mobilenet,a lightweight model,Light-SSD,was proposed by introducing depthwise separable convolution into SSD,which reduced the number of percent of SSD model parameters at the cost of losing some accuracy.(3)Light-SSD can reduce the number of parameters while losing precision.The distillation of Light-SSD was studied by using the method of knowledge distillation.The SSD with high precision and large amount of parameters is selected as the teacher model,and the Light-SSD with low precision and few parameters is selected as the student model.By selecting(38*38*256),(19*19*1024),(10*10*512),(5*5*256),(3*3*256)and(1*1*256),the loss function of teacher-student model is constructed,so that the student model can learn the content of the teacher model.The results show that distillation learning can improve the accuracy of Light-SSD,so that it can obtain the number of small parameters without losing too much accuracy.