Research On Color Classification Method Of Garbage Bottle Based On Deep Learning

At present,the recycling of garbage plastic bottles in China is still based on artificial sorting method with low efficiency,which is harmful to human health due to the harsh working environment.Therefore,the industry has actively carried out the research of automatic sorting equipment for waste bottles based on computer vision.However,the existing equipments have problems of low accuracy and slow,which can’t satisfy the requirements of enterprises.In view of this,in order to improve the accuracy and speed of garbage-bottle color classification,this paper analyzed the research present situations and problems of garbage classification,and researched color classification of garbage bottle based on deep learning.Through the experimental verification and analysis of existing algorithms,this paper proposed the ideas of network optimization,which are compressing model and designing lightweight model.Then verified and analyzed the improved algorithm model.The main contents of this paper are:(1)Researched the basic structure of convolutional neural networks and three characteristics,including local perception,weight sharing and down-sampling.Analyzed and compared in detail the mainstream two-stage and one-stage object detection algorithms based on deep learning.(2)Researched the application of the one-satge YOLO v4 algorithm model in the color classification of garbage bottles.To solve the problem of unbalanced distribution of data categories,the Focal Loss and Label Smoothing smoothing labeling methods are introduced.Model training adopted the freezing training method,and then realized the detection of single/multiple images.The experimental results showed that the m AP obtained 90.7% under this model,and the inference time was 55.55 ms.(3)Researched the application of the compressed model YOLO v4-tiny in the color classification of garbage bottles.The purpose is to reduce the network parameters and improve the detection accuracy by compressing the network model.The experimental results showed that under the same garbage bottle data set,the inference time reduced to 34.48 ms,but the m AP reduced to 89.0%.(4)Researched and designed the lightweight algorithm model Mobile Net v2-YOLO v4.To achieve the purpose of balancing the detection accuracy and speed performance of the model,improved the YOLO v4 network structure.Specifically,the backbone feature extraction network was replaced with Mobile Net v2.At the same time,using Mobile Net for reference,deep separable convolution was introduced.Under the condition that the network depth is not changed,deep separable convolution was used to replace the 3×3 general convolution of the enhanced feature extraction network,which greatly reduced the network parameters.The experimental results showed that under the same garbage bottle data set,the Mobile Net V2-YOLO V4 model has more significant effects,which achieved the fastest inference time of31.25 ms and higher detection accuracy.The m AP was 90.1%,which can better meet the requirements of enterprises for the accuracy and speed of garbage-bottle color classification and recognition.