Research On Garbage Detection And Classification Method Based On Deep Learning

With the improvement of social living standards,the output of household garbage is increasing day by day,and the importance and difficulty of garbage disposal are also increasing.Improving the efficiency of waste sorting and processing is of great significance for maximizing resources and protecting the environment.However,in garbage sorting and recycling,the classification standards are currently too complicated for the general public.The image detection and classification technology can help residents quickly grasp the classification of daily garbage at home,but the current image-based garbage detection and classification methods have the problems of slow detection speed and low accuracy.This paper analyzes the problem of image-based household garbage classification at home and abroad,researches on the difficulty of domestic daily garbage classification,and proposes a mobile-side detection and server-side classification method based on deep learning.The main research work is as follows:In the process of collecting household garbage images on the mobile terminal,in order to deal with the problem of low accuracy and slow detection speed of garbage target positioning due to the many interference factors in the natural environment,this paper proposes Tight Intersection of Union measurement method.Ordinary intersection of union cannot accurately measure the distance relationship between bounding boxes,while Tight Intersection of Union adds the distance between the center points of the bounding boxes as a penalty item to the measurement calculation.In this paper,a garbage detection network is designed based on feature fusion,and Depthwise Over-parameterized Convolution is added to the feature fusion to improve the detection speed.Experiments have proved that the detection accuracy rate of the Tight Intersection of Union is higher than other intersection of union.The garbage detection network proposed in this paper can achieve faster detection speed.In the process of household garbage image compression and transmission from the mobile end to the server,aiming at the problem of insufficient quality perception indicators of traditional image compression algorithms,this paper proposes an image compression method based on generative adversarial networks,which mainly includes improving the measurement of image compression distortion values and a normalization method.In the experiment,this paper compares the compression quality of different compression methods on data sets such as CLIC2020,and the results show that this method can achieve relatively higher compression effects.In the process of household garbage image classification,in order to solve the problem of low accuracy in traditional garbage image classification and ignoring the hierarchical classification relationship of garbage,this paper proposes a multi-task hierarchical classification network based on Efficient Net.Different from independent and parallel garbage classification methods,the multi-task hierarchical classification network realizes the sharing of garbage image characteristics and considers the mutual influence between hierarchical categories.The experimental results show that the garbage classification network proposed in this paper can achieve a higher garbage classification accuracy rate than other methods,and the highest classification accuracy is 97.56%.