Research On Classification,detection And Localization Of Domestic Waste Based On Deep Learning

The effective classification,detection and positioning of domestic waste not only protects the ecological environment,but also realizes the reuse of social resources.At present,the sorting of traditional domestic garbage is mainly carried out manually.Long-term manual labor and in a harsh working environment full of garbage,this method not only causes low sorting efficiency and easy misclassification,but also affects human health.cause damage.With the rapid development of science and technology,the wide application of computer vision,coupled with the continuous increase in environmental protection efforts in my country,the implementation and popularization of environmental protection policies and concepts,these have made the intelligent classification,detection and positioning of garbage.become the trend of future development.Therefore,this paper uses the application of deep learning in computer vision to study the classification,detection and localization of domestic waste.The main research results are as follows.(1)In the classification task algorithm,for the data set of domestic waste classification,the appropriate algorithm is selected as the research object by comparing different classification algorithms.Through experimental comparison,this paper selects Ghost Net as the research object of domestic waste classification,and on this basis,adopts the parallel connection of low-dimensional to high-dimensional feature extraction surfaces in the Hr Net algorithm,and performs repeated multi-scale feature extraction.The structure is improved to enhance Ghost Net’s ability to extract features from input data.The experimental results show that the improved Ghost Net not only reduces the model parameters by 0.77 MB,but also increases the accuracy by 1.8%.(2)In terms of target detection algorithm,YOLOX-s,which is a feature extraction network replaced by Ghost Net,is selected for the data set of domestic garbage target detection as the research object of target detection in this paper,and the experimental comparison is carried out by improving the feature extraction network Ghost Net of YOLOX-s.Through experiments,it is found that the attention mechanism of Ghost Net and the generation method of redundant feature maps are improved.Compared with the original YOLOX-s,the target detection measurement standard MAP＿5090 is not only improved by 0.74%,but also the model parameters are reduced to the original YOLOX.59.06% of s.(3)In the target detection and localization algorithm,the detected domestic garbage is located.First,the corrected binocular camera is used to collect the images taken by the left and right cameras,and then the SIFT feature points optimized in this paper are matched in the area of the left and right images according to the detection target.The matching feature points are clustered,and finally the specific positioning of the target is realized according to the cluster center point and the parallax formula.The experimental results show that the distance error between the measured distance of the domestic waste target and the actual distance is mostly concentrated around 3.5%.