Primary Study On The Image Recognition Of Household Garbage Based On Deep Learning

This paper has mainly carried out relevant research on the identification of resident household garbage image.As an inevitable problem in the development of human society,garbage disposal is an extremely large amount of household garbage in China with the economic development and social progress.Garbage classification has been proved to be one of the effective means to prevent garbage pollution.How to effectively realize garbage classification and recycling,so as to improve the quality of living environment,is one of the urgent issues of common concern at present.Image recognition technology has wide application potential for garbage classification in terms of fast identification of garbage types,detection and classification of garbage,etc.In view of the low efficiency and poor generalization of the traditional classification and identification methods,this paper proposes a deep-learning-based identification method of resident household garbage,aiming to help the classification of resident household garbage and improve the classification efficiency of resident household garbage.The identification model of household garbage in this paper has the characteristics of fast identification speed,simple operation and low economic cost.The identification model can not only serve the mobile phone software for household garbage classification,but also increase the classification efficiency of household garbage in the distribution of residents.It can also be used in the use of artificial intelligence intelligent trash bin and other places,improve its identification effect,convenient for the majority of residents of the daily waste.It can also be widely used in many domestic garbage treatment plants to effectively serve the related business of garbage treatment plants,increase the efficiency of identification and classification of domestic garbage and detection and treatment,reduce the cost of garbage classification,and play a role in energy conservation and environmental protection.The main research contents of this paper are as follows:(1)Because the household rubbish mature image of public and very little data set,in order to better study spam image recognition problem,and comment on our proposed based on the transfer of learning method,we according to the guidelines for the municipal solid waste classification in anhui province,the selection of some of the common residents living garbage,collected through independent film and network formed a HGI-30 residents living garbage data set.It contains 30 small categories of household garbage and 6000 RBG images in total,each with relatively complex background,posture and illumination.At the same time,we also used cropping,flipping and other methods to enhance the image data.(2)Based on Vgg16,Inception V3 and Res Net50,a method of image classification and recognition of household garbage based on migration learning was proposed.Two training model methods,direct training and transfer learning,were used respectively.On the basis of retaining the generalization ability of source model,adaptive fine-tuning was carried out for the sample set of garbage data set.The experimental results show that the model of transfer learning training is superior to the model of direct training in recognition accuracy.We also designed different freezing layers to study their influence on the model recognition accuracy.By adjusting the fine-tuning strategy for many times,the experimental results show that the trainable parameters can achieve better performance when the proportion of trainable parameters is 75% of the total parameters.We achieved the best result of verification set accuracy of 95.1% by tweaking Inception V3.(3)An improved SSD300 resident garbage image detection algorithm is proposed.To advance training SSD300 as source model,additional convolution reconstruction after VGG16 structure in the source model layer,draw lessons from the Inception model structure,we introduce three kinds of hollow convolution layer(3 * 3 conv,rate = 1)(3 * 3 conv,rate = 3)(3 * 3 conv,rate = 5),given the depth of the convolution of the commonly used image detection algorithm in small object recognition effect is slightly less,designed for small garbage SGDB structure of object detection,In this structure,multi-size convolution kernel is used to replace the original single 3×3 convolution kernel,and sparse convolution kernel is added to improve the feature information extraction ability of small garbage targets,without increasing the complexity of the model too much,and thus improving the recognition rate of small household garbage.